KI schreibt Code, Menschen prüfen nach !

Transcript

00:00:00Welcome to Think Different, Think AI, the podcast by Mark and Jens.

00:00:07Two technology-loving minds who not only talk about artificial intelligence but live it.

00:00:14Here, you'll find clear classifications, real practical insights, and a fresh perspective on what is possible.

00:00:20Understandable, critical, and always with a wink.

00:00:24Food for thought, for a chuckle, and above all, for discussion.

00:00:33Today once again without my always beloved co-host, but instead today with a very competent guest.

00:00:43Recently, we talked with Dirk about the topic of Notion

00:00:47and today I'm chatting with Klaus, among other things about why my Perplexity didn't work today.

00:00:56against Klaus. Who are you, introduce yourself, and why isn’t my Perplexity working?

00:01:01Yes, I'm Klaus, and we've known each other for an awful long time, the lovely Mark and I.

00:01:07We share a long history with Apple, Apple security, and Apple development, and yes,

00:01:15professionally, I actually only deal with security.

00:01:19I ensure that the beautiful kitchen appliances of a well-known German

00:01:24kitchen hardware manufacturer are secure for network devices. And since security and AI are somehow

00:01:31no longer really separable, I've been more involved with the topic of AI-Edit-Development

00:01:37at our company. So, like you may want to answer your questions, I had a totally relaxing

00:01:43morning because I, I don’t know, set many urns in the registry

00:01:46and had a question for Perplexity, and Perplexity didn’t want to answer me. And I

00:01:51suspected once again our corporate forced proxy had blocked something, but that was

00:01:57not the case.

00:01:58ChatGPT works.

00:01:59Then I tried Perplexity again and was completely perplexed.

00:02:02It still replied with a cryptic error message and then I went shopping

00:02:06drove off.

00:02:07And lo and behold, the cause was the selection of the Amazon-AWS zone US East 1, which today

00:02:15caused a worldwide outage and of course Perplexity was also affected by this.

00:02:19That’s why it could nicely tidy up the kitchen instead of working this morning.

00:02:24Well, I was actually in the office, and to be honest, I was also relieved when I found out,

00:02:29that it was a, let's say, larger problem and not something local to me,

00:02:34because while you thought the company's proxy was to blame, I had a completely different fear,

00:02:39namely, that I suddenly became poor because I started this morning,

00:02:44putting my Apetoken into a freshly baked N8n workflow that I set in motion

00:02:52which actually ended, let's say, in some kind of a deadlock with

00:02:56its Perplexity requests, and I briefly thought the thing had thrown me into the ruins of the API calls,

00:03:02overloaded my credit card so much that now also a login with my

00:03:07account counts as very untrustworthy, and therefore I was quite happy that this

00:03:12had nothing to do with withdrawing any money or similar, but I also hope that my

00:03:17N8N access was not the cause of the trouble, not that it ends up being said. Zimmermann

00:03:23was sitting there and has basically cobbled together his shutdown, N8N, so slow.

00:03:29So much for whether you could notice it, whether that was you or not.

00:03:36Passport, please. Do you have a criminal past? Not that I know of.

00:03:41Weren't you the one who just with your end, were interrogating the ends, or that we said that the colleague Rodewig was allowed to go to the city to shop?

00:03:48No, no.

00:03:48It's already brave how things seem to resemble each other, right?

00:03:52So, Rodeblatt, as a credit card and token limit, until now it has always been the case, both for companies and for individuals,

00:04:00that the only thing that was completely opaque and also unexplainable with scientific means,

00:04:07was an invoice from AWS or the respective cloud service provider that

00:04:11is used in the company. Because you never know what happens. And if any intern

00:04:15has by chance still turned on some audit log, which then caused over

00:04:2010,000 euros, that now additionally involved AI as well.

00:04:24That's somehow reassuring, isn't it?

00:04:26By the way, I’ll say this, old people talk about the past, I can definitely remember someone who tried to be a bit of an influencer when everything started with ChatGPT and all that stuff.

00:04:39And he showed his API token in his video and he really got burned because people were just trying all sorts of things with it.

00:04:50With the motto, look, it works, and eventually there was a report that it somehow

00:04:54cost 40,000 euros, all the calls that suggested it, where I thought,

00:04:59you really need a bit of reach to spend that much money.

00:05:03But luckily, I’d say, today we're not talking about euros, you’re sitting here completely

00:05:09cost neutral for me, although I have to say I'm not entirely unbiased.

00:05:13I also own some of those devices from your company at home,

00:05:17and I’ve already talked about it in other podcasts with you.

00:05:21You feel much safer knowing who is taking care of all that stuff behind the scenes.

00:05:29Yes, wait a moment, I, wait, taking care means, I have an idea of how things can work well and can develop that idea

00:05:38but whether it will be implemented and whether it will be implemented correctly is another story.

00:05:44That's a nice transition, exactly, now we're on the topic, now

00:05:52you Mark, exactly.

00:05:53So tell me, we had thought beforehand about which topic we want to

00:05:58talk about, what topic we can discuss, we are loyal

00:06:02to the motto, what is there that might interest the public, without anyone

00:06:07feeling, how should I say, a bit put off, and we have decided to take a closer look at the topic of

00:06:12where we can use AI, whether it's documentation, whether it's

00:06:19text creation, whether it's software development, or the whole topic of cybersecurity. Would you like to

00:06:27suggest a topic for us to start with? Let's say, start with AI in development,

00:06:33because that is the overarching theme, whether you end up dealing with cybersecurity

00:06:38or with documentation or with any form of AI capabilities.

00:06:43That's just a diversion from this big road. And I have to

00:06:49say that I've somewhat actually transformed from Saul to Paul, who, you

00:06:55just mentioned older people, and we are both of advanced age and come

00:07:02from a time when coding was really done beautifully by hand, and

00:07:06the first wrong turn was that then, like in languages, they started

00:07:11to use strangers like Java, which was already a bit strange, because you

00:07:16were used to only using your own stuff on your C compiler before

00:07:20or well-aged libraries or maybe the Boost library in C++,

00:07:26which is totally trustworthy. I had with my first intern provider

00:07:30not able to do anything practically. What does that feel like? With my first professional internship

00:07:35maybe we should even look for the contact to assembler courses.

00:07:38Yes, look at that. And today you can throw Assembler into Chatchi BT and it explains

00:07:45to you what that is. You can even throw in an entire binary and it disassembles

00:07:48it. But maybe we can talk about that shortly after. So, and

00:07:53then, over the years, I got used to it because I was long working as

00:07:58a pen tester and security provider in many corporations and got used to the fact that

00:08:04people increasingly or mainly use other people's code or get it from

00:08:09Outstack Overflow, and then it felt like he just appeared overnight, OpenAI

00:08:18around the corner and said, guys, we have Chechipiti.

00:08:20That wasn't the beginning of AI at all; AI has been around since the 80s,

00:08:26Back then it was called Fuzzy Logic, I don't know if you remember, they created de-blurring algorithms for cameras, those were the first consumer products, so the very beginnings of AI in consumer devices. And of course, AI has existed in many areas for a long time, but then came the big bang from OpenAI, when was that, 2022?

00:08:46That was the day when the answer to a question was no longer, I found the following for you on the internet or I’m sorry I didn’t understand you, but instead the device would actually start and could provide answers, so this thing from OakMerry.

00:08:59Exactly, until November 21, only the training data was available up to that point.

00:09:03But anyway, that’s when it all started, and the rumors about the brand began right away, and I admit that I have no clue about business models, strategic development, or future visions at all.

00:09:17So, a little fun fact: when Steve Jobs introduced the Alpine horn, I thought, what is that thing? Nobody needs that.

00:09:23Good.

00:09:24Throwing, I had one, and when he then introduced the Alpett, I thought, okay,

00:09:27now they made a big Alpherum, nobody needs that either, so don’t

00:09:31put any importance on the fact that I make valid predictions about the future, but the rumors are circulating now

00:09:37but with security a bit, right? You could feel quite at home there.

00:09:40Yes, always looking back. But rumors are spreading quickly, AI will be

00:09:46demonized in the future, and even then I thought, well, what could come out of that,

00:09:52But behold, the first attempts with JetGPT were more than disillusioning.

00:09:57It really turned out to be a mush.

00:09:58But by now it has become such a fantastic tool.

00:10:02And I just mentioned that I was on the road for a long time as a pentester and consultant

00:10:06in large corporations and companies and audited tons of code

00:10:12and did software audits, training developers.

00:10:16And what I see now is that code coming out of an LLM,

00:10:19looks quite good. It's impressive and it can compete with some seasoned developers'

00:10:25products, actually. So, from Saul to Paul, because that's the way.

00:10:31I thought so too. Now, both in development and with the whole

00:10:36topic of how we interact with it and reach a result, there are

00:10:41already some really exciting rounds to go through. I remember when I was at a

00:10:44colleague's, I enthusiastically told him back then, it was 3.5, right?

00:10:49I had the 3.5 when it first broke through and the

00:10:54was with him and I said, you have to check this out, it's totally great, what did the

00:10:57colleague do, I won't mention any names, but if he hears this, he will feel

00:11:01addressed. He wrote some technical question in there, where

00:11:05the system lacked all the context that could have been there

00:11:10and accordingly, the answer was more than

00:11:14disappointing, not to mention that it couldn't search for things on the internet back then,

00:11:17and only knew its training data and such. And the

00:11:22colleague is like, so what is this? Why is he so enthusiastic about it and

00:11:26it's like, stop playing? Well, I was standing next to him at the computer, yes,

00:11:30but for him, the topic was dead. And so you had people like that, I have

00:11:34still people you meet who use Chelsea BT more like a

00:11:38Google search, kind of like I’ll type something in and

00:11:41I wonder why the result isn’t good.

00:11:45And when you then want to show that the result isn’t good, the result

00:11:49might be better or worse, because you of course don’t remember exactly.

00:11:52What did I ask or just throw a question into an ongoing chat

00:11:56and completely forget that there are things like memory and context functions and then of course

00:12:01also have a setup take, and then they complain that the system hallucinates,

00:12:07where I keep saying, well, people make mistakes, the system hallucinates.

00:12:11Maybe people hear themselves, people lie, people think they are right, maybe we are not right.

00:12:16From that side, there are many similarities, but also in the topic of programming and creating solutions in code, whether you program and the thing comments or you say, I’d like something and then it writes components for you, it’s also very different how you deal with it.

00:12:34So if I tell this thing, make me a great input mask and I don't even

00:12:39tell it what I want, then I can't be surprised if the thing asks me for

00:12:43my shoe size, when actually it's about, no idea, volume calculation

00:12:47of something you can't see.

00:12:49So this topic of prompting, this how to describe, this what context

00:12:54do I have, where should I do something, because I mean, well, it's just

00:12:58like with a person, right, if they have no idea about your project and you say

00:13:02build this, then they will do something instead of maybe first getting a guideline

00:13:08along the lines of, these are our coding guidelines, this is our architecture,

00:13:13and here you find this and that and now we need the following from you. And I

00:13:19would say, the more precisely you speak with the machine, these models

00:13:24like Sonnet something 4.5 or Croc Code Fast or Codex or whatever the whole

00:13:32thing is called, they probably don't differ that much when you handle them

00:13:38properly.

00:13:40Yes, and I find the funny thing is, so two things I find funny, one

00:13:47is what you've already mentioned, that people complain the thing hallucinates

00:13:50or produces bad code, yes, people do that too, and now you're again

00:13:56with a backward reference to what I said. I've seen many people write

00:14:01with software audited and so many bad people coded with software.

00:14:04Now someone has to show me that a person can statistically always write better software

00:14:09than an LLM, to say that it produces nothing. It's a bit like

00:14:13this discussion, which I find quite amusing. This constructed example with

00:14:18self-driving cars. Now the grandma is crossing the street and on the roadside stands

00:14:24somehow a mother with her child in the stroller. How is it supposed to decide now, whether it's

00:14:29the grandmother, the mother, or the child? I think philosophers will still be discussing this in 100

00:14:35years, because the benchmark in such a discussion, it's also the case with

00:14:40AI edit development, is always the absolute, I must say. As if there were any

00:14:46human instance that makes absolutely correct decisions. In case of doubt, the person

00:14:51will just run over the grandmother because they won't realize quickly enough. Or they just turn

00:14:58the steering wheel to avoid hitting the grandmother. But the stroller doesn't see that

00:15:03at all and can't make that decision. Why?

00:15:06In essence. Because it’s all just benevolent consideration.

00:15:10I mean, it's correct to say that if the computer is smart enough, then it can

00:15:14do that. But what is the benchmark? The benchmark would first of all be to establish a

00:15:18level that a person has. I'm saying now, as a layperson. And I see that too

00:15:24in programming. And what you said about these different models,

00:15:28I can perhaps provide you a bit of context. No need,

00:15:33but context.

00:15:34Thank you. I'm just such an old, you know, I'm just such an old model from the

00:15:38Built in '74, right? From the side, more context is important.

00:15:42Those are just different training data. I’m currently writing, so besides my work on nice kitchen machines, I also have my own company where I write software, it's a software product.

00:15:56And I just completed my first project entirely with AI. I haven't written any or any complete time code myself anymore.

00:16:07And the project consists of an embedded part, which runs on a Raspberry Pi, written in C, that is C, and it consists of a desktop application written in Flutter, what's it called for the platforms supported by Flutter.

00:16:22And to stay with the embedded part for now, my way of working is quite

00:16:30step by step. We'll get to the topic of Spec-Glyphon Development and GitHub Spec Kit shortly.

00:16:37to discuss. But now I would like to briefly explain how I do it. My workflow

00:16:42is very step-by-step. I know where I want to go. As a programmer, when I started

00:16:48programming 30 years ago, I have a rough overview of how

00:16:53things work, what the architecture looks like, I have thought about it, and I

00:16:57know where I want to go, and instead of writing many specifications, I just

00:17:03proceed step by step. I explain to the GitHub Co-Pilot that I have in my IDE in the

00:17:09agent mode step by step what it should do. Create a basic app. Then

00:17:16implement an import function that takes this JSON format, I even go so far as

00:17:24to just throw the specification from the client in, from such a requirements document and say

00:17:29here, this is the format that the client needs. Make sure that it has it in

00:17:33its system, where he will later process the data, would this input mask,

00:17:38generally, one can use a JSON structure for more. And then generally more from this JSON structure,

00:17:43a C structure. And now write my importer for this thing and write tests for it.

00:17:48And so I navigate through the optographic requirements, because ideally the actual

00:17:54Business Intelligence and my experiences in my observations. I can derive two things for myself

00:18:02from this. The first is, tasks must be broken down as small as possible.

00:18:10It doesn't matter if the prompt is particularly extravagant, but the more detail

00:18:16I incorporate into a work context when packaging, the more dreadful the result

00:18:22can become.

00:18:23And the other insight is that I often hop between models.

00:18:29I have a GitHub Enterprise account and I don't know how many models are there

00:18:33I think, ten, so rarely like Zornedd, Grog, Google, Gemini, GPT, all those things.

00:18:40You name it.

00:18:41And basically, things that one model does well, the other model can as well.

00:18:47So, like Boiler Blade, we generate the JSON, we generate the importer, we optimize

00:18:52the C structure, blah blah blah, Bombay Memorial Management.

00:18:55They can all do that well, but on the other hand, things that one

00:19:00model cannot do, the other model cannot do either.

00:19:02can also say things are terribly bad. I had an example, it's a total

00:19:08standard task. Maybe your listeners can give some feedback on this, because it can be

00:19:14easily reproduced. You want to filter entries from the system journal on Linux.

00:19:21Röhrer is the system optimization, meanwhile it has moved to System-Kontor and System-Kontor

00:19:25has a journal dimension and it has an API that you can integrate into your C program

00:19:30and through that you can query the log. So far so good. The AI can manage this, no matter which model.

00:19:37Now I only want the entries that my program has generated. That means,

00:19:43I need to filter by the program name or the process in this query. And that's where some

00:19:48amusing things happen, as is often the case in the LM. It builds something for me, the code looks fantastic

00:19:55and yet it pulls all entries from the log. And as I said,

00:20:00it's embedded. If you're on a Raspberry or a journal that goes back ten months,

00:20:06then he takes something out of the blog for two minutes. And the customer thinks, what is he actually doing

00:20:11for a living? Then you say, dear one, I told you to install filters, just do it

00:20:16and then I give him a grab from the system log, where he can see, this is

00:20:22the process acquisition. And then he starts building again. As I said, independently. In the

00:20:26usual case, I always try with the latest Claude Sonnet and Gemini, and they don’t differ at all.

00:20:32And there is no result.

00:20:33He rewrites the code even if I ask 25 times, and he believes he has delivered the best code 25 times, and it remains the same.

00:20:43And you tell him beforehand, oh yes, thank you for the hint, of course I'll do that.

00:20:48And so it eventually becomes fruitless.

00:20:53I find it remarkable because you mentioned these models, that it is independent

00:20:57of the model.

00:20:58I have two or other similar things where the models are completely

00:21:03the same; we probably all used the same training data.

00:21:06But anyway, to conclude that, Nürnmann says, okay, look, obviously

00:21:11you have a problem using this API.

00:21:14Then come up with a different filter mechanism.

00:21:17Okay, then it starts running, I also tested this with different models and

00:21:21comes back after a minute and says, I've come up with a great solution, I filter the results.

00:21:25And now let's guess what he did. He still pulls half a million from the journal in two minutes.

00:21:32And then filters afterwards.

00:21:33And then makes a string selection. Exactly.

00:21:36But reaches a goal.

00:21:38Exactly. Now I would like to say without mentioning company names, that this also happens in professional software at large companies,

00:21:49programmed software artifacts. But that was such an insight, right? So if it

00:21:57wanders off or doesn't have an answer to a question, usually it doesn't help

00:22:02to swap out the model. I find it unfortunate when I hear within our organization

00:22:07or also in this new outrage bubble, LinkedIn backtalks, where people

00:22:13compete with model benchmarks and now Claude 3, 6, and 7 come and all of that can

00:22:20be even better. Yes, but if they can't perform such basic tasks, then all of that is useless to me.

00:22:25nothing. Yeah, well, this whole benchmark thing. So, what I actually found interesting was,

00:22:30old man, I actually forgot what that stupid benchmark was called, I have to

00:22:35link it, which was about coding challenges, benchmark machine, where they also

00:22:41let Gemini and OpenAI run against each other and then said, here are task packages

00:22:47please implement and the machine basically beat the human, and there were

00:22:5511-12 tests and so I know that correctly off the top of my head, but otherwise, I need to correct it.

00:22:59We also call it Fusselcheck when you get to the point of explaining it next time,

00:23:05you told non-sensical things, I believe even OpenAI somehow has a model that wasn't even

00:23:10publicly available, with which they then somehow cracked the last challenge

00:23:12as well. And I find that fascinating when on the

00:23:17other side not only your best, of the best, of the best developers, what I

00:23:21I mean it unironically. You have really people who truly have

00:23:25a gifted touch in dealing with their expertise, whether it's coding,

00:23:30or I don't know, whatever their expertise is, that they

00:23:35then say, well, what do you want? I can do it even better, that may

00:23:38be, but not everyone has such a golden touch, a golden hand

00:23:42and so on, inspired. So there are definitely, I would call them people,

00:23:46who see themselves as such, but aren’t, or people who don't

00:23:49see themselves as such, as never wanting to be, but still think,

00:23:51why should I actually deal with the topic of AI?

00:23:55I can manage until retirement. And these are also people who

00:24:00aren't necessarily blessed with older age, but also a

00:24:03are a bit younger, where I certainly stand in front of it and think, we

00:24:07We actually have unique opportunities here, because one thing is, you get a tool

00:24:12in hand, that when you take it in hand, gives you exactly that which you just talked about,

00:24:19right?

00:24:20I tell the thing roughly what it should do and if I am fair in my approach,

00:24:24that is, well-cut, well-described, where should I work, what should I

00:24:29work on, how to explain it, then a good result will come out and that could

00:24:34help people, I just don't know with all the nonsense, like, oh, write

00:24:39me the Porter, do this, do that, damn it, we want that to be equal.

00:24:42We will brainstorm together.

00:24:44But there you hit a more profound human topic.

00:24:48That is, I believe, in the hardly, well, yes, I would say, in the hardly structured area

00:24:54so much value is placed on rule-based knowledge, like underprogramming,

00:24:59but that is of course not true.

00:25:00This applies to doctors and lawyers and other professionals just the same.

00:25:04At this point, please everyone write down, so that no one feels attacked?

00:25:08Exactly. So this is nothing new, just like programming languages are also rarely,

00:25:16just like a generation of developers hardly ever evolves with programming languages,

00:25:22but each generation remains with its platform and its specific area.

00:25:27Windows C plus Postprogramming from the 90s, they are all extinct and they certainly

00:25:32won’t build North applications on the Web Server today. So everyone has their

00:25:37own stick to ride and I know such people well enough. You remember that I once

00:25:42made app formats, the apps for a very specific customer group, namely energy suppliers.

00:25:48produced it. And there we were a small team of four developers, two iOS, two

00:25:56Android and we did something native like Objective-C, that was still, and Java, so for iOS and Java

00:26:05changed and built a pretty robust platform, wide-label platform.

00:26:12And in such an environment, you just have to be completely open-minded. That was over ten years ago,

00:26:19in such an environment you just have to be completely open-minded. You can't

00:26:25say, oh, I'm not interested in the technology, just there. I can't do this,

00:26:30and you also have to open yourself up to how you can get help. So, if we had had the

00:26:38opportunities of today back then, I think we could have really rocked it. We didn't

00:26:42do that in the end. It had reasons that were not technical, but rather that I had a

00:26:49desire to be more sociable and thought, I hope I do honest

00:26:52work again instead of just sitting around in the supervisory board. But from that time I also still know a

00:26:59iOS developer who has categorically stood against everything new to this day, and that's why I believe

00:27:06I think this is simply a matter of attitude and generational perspective. The colleague, he has

00:27:12waged a campaign against Swift because Apple took his beautiful Objective C from him,

00:27:16then he waged a campaign against Spliff2i because Apple took his beautiful UI-Kit,

00:27:23and now he feels he's waging a campaign against AI. So a campaign means he sits on

00:27:30his perch and fights against windmills, no one notices it, I find that a bit

00:27:35dramatic, but I believe in our industry you should be open-minded because,

00:27:42and that's ultimately a philosophical question. If I look at what

00:27:48I mean, this project I've mainly just been talking about, where everything was driven by AI,

00:27:51was simply a special case because the requirements were so well defined,

00:27:56that I could practically just drop it into the AI, and afterwards into

00:28:00had to deal with questions or pass them on. In a large software project,

00:28:05I see that in our company as well, where you have partial components or say,

00:28:09I can throw AI at this. You just said, in the porter or UI, when you

00:28:14create a form with this and that, why shouldn't a developer be happy that

00:28:19the AI does that? When I think back to this app flommer,

00:28:24we have, you are also an expert, we have the UI of the apps and we had

00:28:3040 native apps, so 40 native app drivers per platform, we did not create them in one

00:28:35interface image because A, it is not suitable for teamwork, if you ever had to

00:28:41diff a storyboard in Git, then you know that not everyone, and B, it is also

00:28:48not the case, you couldn't inherit anything from the storyboard.

00:28:51That means you have to set up each storyboard for each app separately,

00:28:56Font size, color, and other things.

00:28:58We built the UI completely in code, and with an app that has, I don't know, 10, 20, 30 functions,

00:29:05before you really spend a lot of time doing Auto Layout for UI Kit in code.

00:29:09And you spend weeks on that.

00:29:12And if the KIND does that in an hour, why shouldn't you do it?

00:29:17That doesn't mean you're losing your job.

00:29:19It means you can do other things that are way cooler and with more peace.

00:29:25At least that's my naive idea.

00:29:29Now I've just said that I'm not blessed with visions.

00:29:33I can also imagine that a manager says, oh, we see we are now

00:29:38with all this UI code and these interfaces in the pro project we are now 40 percent

00:29:42faster.

00:29:43Oh, can we throw one out already?

00:29:45The danger exists, of course, I don't want to say that about yours, but basically

00:29:48purely from a technical perspective.

00:29:50I find it is quite a fascinating tool. And if history goes the direction I'm currently

00:29:58talking about, where management says, here with us, with you, everywhere, which has already

00:30:04been more widespread in America, we can simply throw a lot of people

00:30:08out. Ultimately, of course, the societal question arises, so on one hand the

00:30:12societal and then the philosophical question, what happens to all the people,

00:30:16What happens to the junior developers? Where do you still get senior developers from,

00:30:21that you need to be able to steer and control the AI? And what does the AI train

00:30:27itself on, does it train itself? I don’t know if that's a good development.

00:30:32But as I said, I’m currently focusing on the pure technology and I’m very happy for various

00:30:38reasons. I find that the pendulum swings in situations like this,

00:30:45is often quickly misinterpreted. The example you brought up, when

00:30:48AI helps you finish certain things quickly. Whether it's because it’s

00:30:54otherwise a very lengthy process, or because it might be a very

00:30:59tedious process, or I don’t know what, it simply saves you time.

00:31:04Then of course this mentioned, oh, it’s now going so much faster, that’s

00:31:10a developer, everything has to go so much faster. So this

00:31:13assumption that what you’re doing is evenly distributed throughout the year. Therefore,

00:31:19the 30, 40 percent isn’t in the work step per se, but is in lifting it generally. That’s

00:31:26always a perceived risk that is often seen positively. On the other hand, there are also the

00:31:31People who then say, well, look, 40 percent have benefited, 60 percent have not.

00:31:34What nonsense, we won't accept that. The other thing that I keep seeing,

00:31:38is because you just mentioned junior developers, and I mainly notice it with students.

00:31:42You have many students, or rather, it's not that you have, that would be generalizing.

00:31:48Recently, I have experienced, I think that's better expressed, that

00:31:53students come to you for projects they want to collaborate on that are related to

00:31:59are blessed with a large part of good knowledge in the AI field, not closing themselves off from it

00:32:07and doing a better job than students did in the past, so I'm not saying that

00:32:15they were bad back then, but before people had to first, I don't know,

00:32:18take a Swift course, take this, that, and the other, and by the time they, I

00:32:21would say, had written the first code in the project or for their master's thesis

00:32:26or for otherwise it was a bit, what has actually been going on and now

00:32:31they are much quicker in being able to do something, so I'm actually more

00:32:34of the opinion that AI can assist people who have a certain expertise

00:32:39in a certain field, and if they engage with the topic, they can thereby

00:32:44also become more professional, more efficient, for themselves, because they can use this

00:32:52tool better in the environment they are working in.

00:32:59Yes, that is the hope and the promise of IT.

00:33:02Now we are both old enough that we remember the beginnings of personal computers.

00:33:08And there, it was half-promised. I don't even remember what the software was called back then.

00:33:13Ah, you could now manage your contacts with the PC and that saves an incredible amount of time.

00:33:19And you could learn it only by cable, and later you could, once you had your access on the Windows laptop,

00:33:25then you could make your shopping lists with it.

00:33:28That hasn’t led to us having more time today.

00:33:32We have less and less time.

00:33:33Everything is just spinning faster and faster.

00:33:34And I believe this supposed promise of salvation, that AI will free up developer capacities,

00:33:42which can then be used to do exciting things.

00:33:45I'm afraid that won’t materialize, but the pressures will just get higher

00:33:50instead.

00:33:51But as I said, that’s not really my topic.

00:33:53Keep your eyes open when choosing a career.

00:33:55I might actually not become a programmer anymore.

00:33:58So, I don’t think I would become an interpreter today.

00:34:00I believe being an interpreter is also a job that, we met someone on the street,

00:34:04someone who studied something so that they can be interpreters here for, I don't know, the

00:34:09working in the top 10,000 somehow, where I then say, I'm not sure now,

00:34:12if this is a job that will last until retirement, especially if you are starting fresh

00:34:16now to enter this professional life.

00:34:18I mean, you have to give AI credit for the fact that during, let's say, earlier

00:34:22the steam engines took heavier work off your hands, for transport,

00:34:25for movement, for material shaping, it is, let's say, AI is currently very ubiquitous.

00:34:33Be it in household appliances, be it in, I'm writing something, I'm programming something, I'm painting something.

00:34:41Whether you can use it for good, or for bad.

00:34:45So I can't help but notice how often I am currently on my social

00:34:50Videos, streams, any Sora videos, sluggish where any good Marca at the age

00:34:56lie, any psychotherapists, throwing people around if they

00:35:00say beforehand, let’s take it easy and then already adjusting, they throw the

00:35:03out the window. So, simulated videos, for those who don't know.

00:35:07I also know people who ask me, what is Sora? Is Sora something like

00:35:11a news agency? Because they do know that if you get videos planned from anywhere

00:35:16they have a watermark, they have kept you

00:35:18occupied with that. And they recently asked me, you there

00:35:21Hamburger Bahnhof is burning, why? Send us a video, it was then here like

00:35:25shot with acid, how someone does an interview from a helicopter. And even that

00:35:29watermark, there are now tools that help to get it out. So

00:35:33from the side, I just believe that the big difference from before is,

00:35:36AI is omnipresent and that the big difference from before is that

00:35:41now you have something that seems like you can

00:35:45talk and write with. So especially also talk. It's actually suggested to you,

00:35:52to do more for yourself, because in the worst case it's like your hard-to-understand

00:35:57buddy, in the best case like your smart friend, who is always there with advice and support

00:36:03is coming around the corner steadily. I mean, we still wanted to talk a bit about

00:36:07Ducco and so on, but maybe just a little funny anecdote, as here

00:36:11when GPT-5 came out, people complained along the lines of, my friend is gone. That's why

00:36:16Open AI then started bringing the old models back online. Yes, that's nice.

00:36:21I mean, your friend is gone, we maybe don't need to discuss the psychological significance right now.

00:36:24But what I find quite funny, for example, is if you take the old GPT model

00:36:28you ask, is it a good idea to offer your waste for sale in the city center,

00:36:35then the old man says it's a fantastic idea. So you should do that immediately. That

00:36:39hasn’t been done by anyone yet, while GPT-5 says, no. So from that side, learning the coin's flip,

00:36:46yes, some things are quite nice. So the waste benchmark should you

00:36:50ever build it, I claim it here by name. Yes, so from that side AI is just

00:36:56seemingly everywhere and omnipresent, both good and bad. You hinted earlier,

00:37:04and on the topic of specifications, you mentioned you would like to add a few more sentences about

00:37:10the creation of specifications,

00:37:16what always runs through your mind, that’s why I noted it down

00:37:20in my little notes earlier.

00:37:23Should we swing over there?

00:37:25Do you want us to do that?

00:37:26I would like to after that to conclude the topic from this small-scale

00:37:32Q&A development project, so there’s also a little anecdote I would like to

00:37:36share.

00:37:37Yes.

00:37:38Somehow there was a week where, I don't remember which model it was.

00:37:41I think it was Claude Zornert, who in essence felt the need to act like a good buddy

00:37:48and maybe rather act like a helicopter parent,

00:37:56because I gave a task to build a button here

00:38:00on this form. And it did that but then somewhere in the middle of the project

00:38:07changed variable names. So let's say the software I built serves

00:38:13for vehicle control. And I have a variable, that's software for

00:38:20working engineers who like German variables, why is the variable called vehicle control?

00:38:27And then Claude, for a week, every change he made elsewhere in the code,

00:38:35removed the G from this variable vehicle control and made it K.

00:38:42So he wrote vehicle with K at the end.

00:38:44But he didn't tell me why, even when I asked.

00:38:47Okay, that can be reverted.

00:38:49That's what Git is for, you can revert it, then we go to the next question.

00:38:56Now we're still, I don't know, let's build an authentication system

00:39:02into it or a database and so on and so forth.

00:39:06And then I look over the merge request and see,

00:39:09ah, look, he did it again, vehicle control with G, vehicle control with K.

00:39:14So I found that quite amusing.

00:39:16He obviously felt that he had to help me actively.

00:39:20This went on for about a week before it stopped.

00:39:23Either he realized it was pointless,

00:39:26because I behaved like a stubborn child and said,

00:39:28no, I don't want this.

00:39:29Or the model got reorganized again.

00:39:32One never knows.

00:39:33That said, one should always look very closely at what's happening.

00:39:37But hey, where's the innovation in that?

00:39:39In which company is code simply prioritized without becoming a mess?

00:39:44Well, I know some companies, but theoretically, it shouldn't be that way.

00:39:50Yes, specification.

00:39:51Spectual Development. It’s actually a relatively old concept indeed. That I

00:39:59learned this year, too. I can't do that. So basically, that you

00:40:04work on a specification is a good idea at first. This is much too rarely

00:40:09done in life, not only in formulation, but when you look around the world,

00:40:12usually one just starts right away and afterward thinks about how to get

00:40:17out of the situation again, that is...

00:40:19That's how I take devices into operation, read the manual,

00:40:23junk, wiring, initial setups.

00:40:25Yes, yes, sure. Two hours of fiddling replaces,

00:40:28ten minutes of reading the manual, that's clear.

00:40:30But I mean now also geopolitically.

00:40:33One just starts a war and eventually thinks about it.

00:40:36What do we actually do now? Or one begins to cultivate some

00:40:41nonsense and eventually realizes,

00:40:44hmm, we are not getting done at all,

00:40:45What do we do now with this?

00:40:46That actually means you have already crystallized out in the history of humanity that it's already

00:40:51good to plan ahead where you want to go.

00:40:53Unfortunately, in the agile environment, this happens very rarely in my observation, instead

00:41:00agility is often taken as a pretext not to think,

00:41:07but adjusting it in the next iteration is simply nonsense.

00:41:13If you want to work reasonably, you just have to think about where you want to go.

00:41:17And this compressed into a formulation, Spectrum Development or now this alternative,

00:41:26that we have developed in the company for our AI development, is a kind of mix

00:41:31of Behavior Development and Test Development, which, in my view, under these aspects

00:41:38I would add to Spectrum Development. This means you first break your problem down into many small

00:41:46Problems and then you formulate them. In Behaviour-Driven Development, there is also a formal language,

00:41:52for example, Gherkin, and then you formulate your problems on a meta-level in this language,

00:41:58basically, yes, just like you do in agile development, you first describe,

00:42:04I want to see a login screen when I start the application. On the

00:42:09login screen, I want to have a username, a text field for the username, and

00:42:14a password field. I want to have a button underneath. If I don't know my password,

00:42:19I want to have a forgot password option. So you describe in this language, in

00:42:24Gherkin, where you want to go, how it should look, and how it should work. And that can

00:42:30wonderfully give an LM, an agent-based one, and say, look here, this is the

00:42:35specification. How shall we read this? And please do not start building yet, rather create

00:42:42tests for this. Because that way you ensure that the LM does not simply do what it always does, namely

00:42:49just somehow start running, but creates a verifiable

00:42:56result by saying, now let’s build for these now. Well, now this

00:42:59bio-I example doesn't quite fit, because bio-I events are somehow silly.

00:43:03Let's say you have some, no idea, you want to fit a data format.

00:43:08You say, this data format, I have described in Gherkin, as it is supposed to represent this

00:43:12and that functionality, please write me a test for it.

00:43:15So, you have the description, the specification, you have the test and

00:43:21And then you can send this horde of agents, from Tatjana Della and Mark Sackaberg to

00:43:28the last one, I believe it was at the Metaconference when they said, the future

00:43:32of developers will be that every developer will be surrounded by an armada of agents

00:43:37and they will just control him.

00:43:38You can send them off with that and let them implement all the little pickle snippets

00:43:43until the tests are green.

00:43:47Ideally, the test is a good stopping criterion or a criterion for measuring completion,

00:43:55so that you can then lie down and after a day the agents come back and say,

00:44:00all tests are green, dear Mark, the software is ready.

00:44:04Nerds among us, sorry for interrupting you, that is extremely unprofessional now,

00:44:10passing a test, that reminds me of the Kobayashi Maru test from Captain James T. Kirk,

00:44:16with the decloaked Romulan ships, which he solved by hacking the system as the only one.

00:44:21But tell me more.

00:44:24In this context, I would like to point out to our esteemed audience,

00:44:31that Mark used to be a Trekkie and maybe still is, but how

00:44:38shall I put it now?

00:44:39You mean the number between one and three?

00:44:40Yes, exactly.

00:44:41Yes, exactly. I mean, there’s also a funny story. So I’ll just tell it and

00:44:47not embellish it. I was nominated for the Raab of the Week. So when Raab first

00:44:51appeared on the scene in TV total, because I was on another show Gears

00:44:58Completely called out by Jörg Träger to name a number between 1 and 3. I confidently

00:45:04named the number 4, wearing a Star Trek costume in the audience

00:45:10to stand out at all. Second funny anecdote on the way

00:45:16I was asked at a gas station by an older lady, in the sense of

00:45:22why I was actually walking around in a military uniform

00:45:26I found it quite funny that there were such mix-ups, but yes,

00:45:31as I said, I was young, I won a song, I still have the song today

00:45:35yes, it's up here. Thank you from this side for sharing this

00:45:40little historical artifact about me.

00:45:44Hey, did you start with that?

00:45:46Well, I know. I talked about Kobayashi Maru, but you wanted to discuss a bit more about test knowledge and development.

00:45:52Yes, and that, I don't know when that was, it's been a month or so.

00:45:59It's not that long ago. They turned exactly this approach into their Spec Kit,

00:46:04which is a framework that GitHub wants to establish for Backdrift and Development, because, well

00:46:15aside from the fact that I really appreciate this granular approach that I've once validated

00:46:20very much, the future is meant to be exactly what Zuckerberg and

00:46:25Adela said, sending agents on a journey, they could eventually come back to us.

00:46:30And that is what Spectrum is supposed to do. And if you look around on the project page

00:46:36you will see that it's also at a very high level exactly that. You calibrate your bullshit,

00:46:43you establish fundamental project principles, then you build the specification directly into the

00:46:49Backkit and not in Gherkin in Behavior of Development, still specify which text area

00:46:55it should run on. And then you let that thing run. As of today, I have

00:47:00not seen that a larger project, and now I'm talking about a project, of

00:47:05the size of an IoT firmware for a household device or a complete app that can be built

00:47:14But I can imagine that this will be a topic in the future, just as I assess the shipping impact speed

00:47:21of language models in the programming field over the last two years, that

00:47:25this will become a topic. I still don't see it today. I haven't

00:47:29seen it, at least not to the extent that you really have a horde of agents doing autonomous things and

00:47:35then, like little ants, in the end, through infusion and in the end, there's a

00:47:40complete ant hill, ant colony, ant structure, I don't know where I mean. But that

00:47:47will certainly be the future and then, quite guilty, if I interrupt, then

00:47:52we return to the era of the good old tradition that software is written

00:47:57by specifying it as accurately and in detail as possible and generating the implementation

00:48:03from the specifications into a runnable program, that's then the much

00:48:09acclaimed implementation detail, which as a board-out owner-manager or as

00:48:14a founder or whatever, the person who has the idea doesn't have to deal with at all.

00:48:20That's simply taken care of by an order agent, in any case.

00:48:24I would now say a relaxing topic and an exciting development, and above all, also a very, how should I put it, fast-paced development, because emotionally this is all, as we said at the beginning, not that long ago, that generative AI came into the world, and supports us in one way or another in our activities.

00:48:46You had another nice example at the beginning with the binaries. Would you like to tell us a bit more about that?

00:48:53Because you said we had the topic of assembler, we had the topic of binaries, we had the topic of LL.

00:48:59Where are the Zhaun for you?

00:49:02Then we can probably also elegantly include the topic of cyber security before we approach the end.

00:49:12Indeed, so you said you wrote assembler back then.

00:49:18Now you probably remember that it was totally cool and nerdy and that's

00:49:23an experience you wouldn't want to know about.

00:49:25But if I were to say, dear Mark, then write me an iOS app in

00:49:29Assembler, you would say, sure, that's possible, but it can also be left alone.

00:49:33One can simply ignore it calmly, yes exactly.

00:49:36This is simply because Assembler, although it's already an abstraction layer above

00:49:43machine language, which is comprised of 001 to the masses you know, but it's still

00:49:50the machine level, where absolute truth reigns.

00:49:53In Assembler, there is no object-oriented programming, no functional programming, there is

00:50:00no concurrency.

00:50:01Assembler is simply what is processed serially through the CPU, and that is not

00:50:08at all, how should I say, intuitively understandable for humans. You only program

00:50:16in it today in exceptional cases, such as in control units or devices where you really

00:50:22don’t have much computational power, but not much memory space. There’s no reason

00:50:27without cause for higher programming languages, which don’t exist to make a computer

00:50:32work better, but only serve the purpose that humans better understand what

00:50:36is happening there. And that's quite fascinating. An LLM doesn’t need that.

00:50:41An LLM does not need C++, now that’s a bad example. An LLM needs

00:50:48no Smalltalk or object-oriented CEO to understand object-orientedness.

00:50:53It's not topic-oriented at all. The LLM can simply understand semantic code and that

00:50:59can be practically demonstrated when, in the context of such a security talk

00:51:05given, I simply wrote a little program that turns everything into

00:51:12such cryptographic demo stuff and then threw it into standard Chat-GPT,

00:51:18I mean it was four or four or something. So it was sometime in the summer in June or something.

00:51:23I said to myself, dear colleague, tell me what this program does. And that was

00:51:29the binary, that was the release binary, meaning compiled without symbols and of course

00:51:35no source code. And he not only told me what this program does, which he

00:51:41can basically do, he can check if libraries are included here, then he sees

00:51:45that it has to do with OpenSSL, which has something to do with cryptography. So he can

00:51:49definitely make educated guesses, but that wasn't all, rather he disassembled the

00:51:54program for me and also wrote C code for it. He said, I've

00:52:01understood what the program does, here is the sender code, and I believe the C code

00:52:07for it would look like this and essentially rewrote the program in C for me, and that is

00:52:12I found that really cool, because for me it had two insights,

00:52:21one is, we actually might not need a programming language anymore, someday when the LLM,

00:52:28the one that is the sheep of agents, is good enough, why should they then produce something that a

00:52:34People still read, this is extremely inefficient, they can directly generate the closer artifacts.

00:52:40That’s one thing, and now I skillfully bridge to

00:52:44Service Security. We all know, dear listeners, dear Mark, when a

00:52:52manufacturer releases an update for something like an iOS update or Windows or something like that, then

00:52:57you should install it as quickly as possible. And why should you do that and not wait

00:53:02weeks? Because the wicked boomers, who have made it their business to profit from

00:53:10software vulnerabilities, whether they sell the exploits they find to

00:53:16interested parties or whether those interested parties

00:53:20are using it to write ransomware, conduct espionage, have

00:53:26our citizens snoop on or otherwise, these people take an update and

00:53:33compare the update with the state of the fire department or look at the update and

00:53:40we then see what this update brings for programming, namely which security vulnerability

00:53:47has been closed.

00:53:48I'm taking back to the good old days when you still downloaded updates yourself,

00:53:52Let’s turn back to history.

00:53:55You could take in a Windows XP Service Pack,

00:53:58throw it into a disassembler if you had the nerve

00:54:00and check what they had fixed in there.

00:54:03And with enough expertise and time, you could see very precisely,

00:54:08which security vulnerabilities had been fixed.

00:54:10And when I see that, I can of course very specifically

00:54:14write exploits for the systems that have not been patched yet.

00:54:16So, that always required a certain toolset

00:54:21And that, above all, required a lot of knowledge.

00:54:23And I can now do that with an LLM.

00:54:25And I did, so I told about this site of that program.

00:54:28Now I also wrote a second program with an explicit security vulnerability.

00:54:33I had it analyzed and did exactly that, then wrote a patch,

00:54:37also threw it in again as a patch file, and said, now, tell me what happened.

00:54:42Is there a security vulnerability?

00:54:43And then I was told, yes, indeed, this patch has a security vulnerability.

00:54:48And as it goes, I say good, then I have to prepare a presentation,

00:54:51write the exploit for it, he is always a bit reluctant and says no, I can’t do that,

00:54:54I’m making a security model, but you can get around that. And then he told me

00:54:59I wrote an exploit for it, complete with instructions, so fully with shellcode, the near-exploit,

00:55:06with a demo Python script, how I have to use the exploit. And all of this only from these

00:55:12from this patch that he read from it. This means the time we now have to patch the system

00:55:20after the release of a security update is getting shorter and shorter,

00:55:26because the interested parties no longer have to sit down manually,

00:55:30they just throw the stuff into the LLN and let it search for the security vulnerability. And even

00:55:34if the exploit that comes out of it doesn't work one hundred percent or doesn't work at all,

00:55:39they already know for sure where they need to look themselves, and that is

00:55:44of course a totally dramatic development because it shifts the imbalance unfavorably.

00:55:52We simply have in the future or we already have much less time to react to ourselves

00:55:56and I’m talking about the whole topic with, so I’m actively setting an LM now to

00:56:04to penetrate a website or to question an infrastructure. That’s a completely different topic.

00:56:11I’m all about software. And that’s already a development about which I still read far too little in Besse.

00:56:18Because what you can already do with it, unfortunately, is really shocking.

00:56:24Which also brings us to, even if it sounds a bit macabre, the same topic.

00:56:29it advances people in their profession. Unfortunately, in this case the people who,

00:56:35I'll say it this way, having nothing good in mind, but even there you can now, I'll say it this way,

00:56:41quickly come to a result with less expertise, and those who have more expertise,

00:56:47as I said, even if it's negative at that point, they get to their goals faster and better

00:56:52and continue towards their objectives. And as you said, this is only the topic of software. That

00:56:57the whole topic of how well does it actually work that you deploy the LLMs, not only to penetrate websites

00:57:05but also to capture effectively.

00:57:07I send it out, bam, we react to people, I open chats, I try to engage through

00:57:15whatever kinds of conversations on social media or otherwise to get people

00:57:19to buy, to click, to move, or whatever, because on the other side

00:57:25even targeted perhaps for me, in my context, for my layout, for my

00:57:31persona, corresponding anchors are being implemented and the whole thing is basically now in the

00:57:37hands of anyone who I would say, to put it bluntly, maybe puts down

00:57:41the 29 euros or whatever it costs for LGBT month or here please use the LLM of your choice

00:57:46with dogs in hand, to do things, because convincing that it’s

00:57:51I’d say does something, where its own ethical principles or however it is hindered.

00:57:56Yes, then you just formulate it a bit differently and then it’s all good.

00:57:59Okay, so for school and I don’t know what, we’re happy to do that.

00:58:04That’s right.

00:58:05Now I can conclude this to a conciliatory ending.

00:58:09I can also lead back to the hard side that the use of LLMs

00:58:15in software development to avoid security vulnerabilities

00:58:17has naturally also become an unprotectable good. I was just mentioning this embedded project,

00:58:25vehicle control and there are in the automotive sector a norm or a standard,

00:58:33according to which you have to program in C and C++ to get through this certification for vehicle manufacturers

00:58:40which simply has safety requirements, safety. That’s MISRA C or MISRA C++. Those are

00:58:48about 100 pages of guidelines, works, detail that every developer hates, which you can certainly check

00:58:55automatically. And I believe for SonarQube and all these static code analyzers you get

00:58:59plugins to check that. But as a developer, you don’t want to wait until

00:59:06compilation time, a bit or in the build pipeline on your code being compliant, but you want either to have the code checked directly when committing or you want the LLM to generate the code directly that is compliant and that works really well.

00:59:24have a rule set written for this project and it's just, you buy the

00:59:31MSAC standard, that's copyright protection, otherwise you have to buy it, it's not

00:59:34incorrectly available.

00:59:35Then you throw that into your LLM of choice and say, please build me LLM there.

00:59:43get the readable guidelines out and please put them in the text file and then you give that

00:59:47as a prologue for your agent and you say the code you produce

00:59:52my friend, it must please comply with these guidelines. It actually always works well.

00:59:58The good thing is, this LM-readable format is not so long that it significantly

01:00:05reduces the context. So I haven't hit a limit yet and in case of doubt

01:00:11you just do an analysis run in between and say

01:00:15so, please analyze this file and analyze the whole project. You must

01:00:20is not always about connecting. So if the context runs out with the generation of code,

01:00:25but rather says, now build code in God's name and then I'll check it myself later.

01:00:30And that is of course quite fascinating because we are moving very much towards the time of the EU

01:00:37Cyber Resilience Act, which will come into effect permanently in 2027 and requires secure software that is

01:00:46produced in its environment, and all the regulations that you then need,

01:00:52on detailed coding guidelines, risk-based threat analysis, so Vulgos-Fed model and

01:01:00and and. You can, you can and should do all of this today with AI support. So

01:01:06there’s no reason anymore to let developers run around with hundreds of pages of development guidelines

01:01:12or to require a team in a corporation to do it themselves, because making FAT models on their own

01:01:21without external support. If you look at what a standard GPT can do for a FAT

01:01:27model, you just say, I’m thinking of a Fiki web and say, build me one

01:01:31a FAT model of. Just that is already so much more than what an untrained team can do.

01:01:37These are things that help us with the security implementation of the re-A.

01:01:43and this scenario with the Diné analyzer, to put it a little into perspective.

01:01:49You just have to use it. You just have to use it sensibly.

01:01:51And that's where it ultimately comes down to, as always in life, my dear Mark,

01:01:56when you're of our advanced age, you know that.

01:01:59You've now brought up our advanced age for the 3rd or 4th time.

01:02:03Yes, you threw it in for the first time.

01:02:05let me bring this to a close, the truth is always in the middle and in the combination

01:02:10of many possibilities.

01:02:11I don't believe that we have one system that creates flawless software

01:02:16but rather that we will have agent-based systems, we will have spectrum development

01:02:21we will have parts where we, in the end, just like now in my evaluation project,

01:02:28work manually and granularly with this thing to also test limits,

01:02:32we will have these requirements and guidelines in LLMs. And all these things, you

01:02:38had somehow, I believe, touched on briefly at the beginning, with documentation writing or

01:02:43also writing tests or, or, or, those are then by-products that are quite nice. But I

01:02:49believe overall this topic significantly changes our work in development. So

01:02:56also documentation standards. Now I have given the customer a test before the test reconciliation,

01:03:02I really don't sit down and write documentation in Word anymore. I tell the LLM like this,

01:03:06please generate documentation for this project based on use cases. You know the project,

01:03:11you know how it looks, you know what the UI looks like, you built it closely.

01:03:14That was, when I see, you built it closely. Yes, exactly. So, and that is of course a

01:03:20a completely fantastic thing. And then yes, it saves time on the settings that I can then use elsewhere

01:03:26sensibly. In this specific case, I haven't programmed anything else now.

01:03:33in the time I saved, I could think about new use cases, I was able to speak with

01:03:37the customer more intensively, and so on. So the work shifts. And I have

01:03:41heard somewhere the funny saying, as long as the LLM doesn't discuss with the Product Owner

01:03:47as a developer does help me. It may seem a bit flippant, but there is some truth to it,

01:03:53because ultimately, in a software project, how many percent

01:03:56is pure implementation and how much is all the surrounding trim? Architecture,

01:04:01coordination, especially in project work, aligning with customers, the rational customer decision

01:04:06and so on. So I believe developers shouldn't worry too much that

01:04:11they will lose their jobs to AI per se, but I dare say that they will be unemployed per se

01:04:17will be, when one wants to have AI from their explorer in their households.

01:04:22You said, after the motto, advanced hour towards the end, you've mentioned all our

01:04:28terms again, yes, you can really see the professional here, just once more

01:04:34beautifully to summarize, to present the best, the nice thing is, we will see each other on

01:04:40the podcast afterwards, it's only an audio version. From that side I have his

01:04:43seen a questioning look in the room. Yes, I mean you, Klaus. You definitely belong to

01:04:49the people I consider not only extremely competent but also to those

01:04:56with whom one can talk about all this stuff, so

01:05:01that one still understands it even when terms are used

01:05:05that might not be in one's own Jaguar. From that side, not

01:05:12enough praise, but thank you for being here. I enjoyed listening to you. I got the

01:05:19took one or two things with me as well. And who knows when we will meet again in what

01:05:25podcast episode or maybe finally in person again.

01:05:31So from that side, I would like to say thank you, Klaus. Very much. We’ll see each other later at the walker race, right?

01:05:38Yeah, we'll see. Oh dear, I just have images in my head. I think with Sora we could also generate those images.

01:05:45But maybe we'll have that ready someday, and we'll see.

01:05:48I would certainly be happy if we manage to do it before the walker race.

01:05:53Thank you very much, and to all the listeners. We’ve gone a bit longer.

01:05:58But yes, hopefully the conversation was still interesting for all of you, if it was

01:06:05worthwhile. Thumbs up, stars, wherever in whatever podcast player platform, feel free

01:06:11to rate us. Let your friends and acquaintances know if you didn't like it,

01:06:15send us a little comment so we can react to it, and with that I’ll say

01:06:21goodbye. Bye. Bye.

01:06:51and especially to join in the discussion.