# AI Shifts Bottlenecks to Human Alignment

**Podcast:** Stories Connecting Dots with Markus Andrezak
**Published:** 2026-04-13

## Transcript

Stories Connecting Dots.
A business podcast to discover a world of possibilities through stories told.
We have to do with the topic, about what we're talking about today.
And I saw a few prototypes of you, about things we're talking about today.
And our topic is of AI, what that means for consequences.
Everyone has FOMO without end, they are confused.
The ones don't have it, the others don't have it, the others don't have it in every detail and come from YouTube and the latest news.
and they will go through the weekend and think about it, if they go to the end or something.
And I think you're also someone who is in the danger of being there, to throw it down and everything to see like a game-kind, so a bit like I or so.
Exactly, we're going to talk about today.
And so first of all, thank you for being there and taking the time.
And maybe the first question, who are you, what do you do and what does...
Olli, Olli, before I get into our history, I'm going to get into it.
Thank you and thank you for being here.
It's really something special for me.
Yeah, who am I?
Oliver Schwarz.
I live in Rheinland, I'm 55 years old and I'm already busy with computers.
That's the classic background.
I'm responsible for the development of the middle-aged company in the middle-aged company.
team an software Entwicklern.
We are a little bit of a part of the world.
And that gives me maybe a little bit of the background for why we are today talking.
You were actually at us for four weeks and have a very with a very interesting talk from my opinion.
Also really super, super good.
He helped us very well and was also quite inspiring.
That I have also made as feedback.
And of course...
As software development are we...
Yeah, we surf the AI-Welle, I would say.
Or at least try it.
Ab and to throw it in the water.
But that's at the moment what's also about us.
Yeah, also you have a concrete number.
So I can name you now Jungspund.
I would have never been taught to concrete numbers.
But there is the number, very good.
I'm older, I've won or lost, I've won, I've seen.
But it's about today.
We have a practical practice here on our channel, who has a lot to do with it and builds and works and works on all levels, so that he also as a leader of the team, what it's about to ask questions.
And on that, what's happening at Testo, we're at the end of the podcast, which is probably for many people also relevant.
In my company are the next bottlenecks, which are also coming back to me.
I would like to ask for the story, which we have, what is our story?
We have, I don't know, on Twitch, on Twitch, to Corona-Zeit, I can't say before, I started.
in Corona rein und jetzt vor einem Jahr, glaube ich auch, haben wir noch mal öfters das zusammen gemacht.
Auch schon zum Thema AI damals, da war noch nicht so viel klar wie heute und wir haben noch mehr rumgeraten, was daraus wird, nicht wird und die Tools, die wir heute haben, standen nicht zur Verfügung.
Also Cloud-Code gab es, glaube ich, wirklich noch nicht, als wir es letztes Mal gestreamt haben.
Das ist zu jung dafür.
And we had the time the user perspective, which is now a bit of a spalting.
There are people who use the known Chat-interfaces, Chat-GPT, Claude and so on.
And based on the experience they have been able to do it.
And they say, well, it's cool.
Maybe I can even get my doctor in a conversation, because I already have a pre-analyse.
But it's just a better machine.
And if it's complicated, it's difficult.
Because my context, like you said, is difficult to push.
And the new tools like Cloud Code, and so forth, with Context Engineering, give us a different chance, a system to build up and to exactly what we need.
And ups, surprise, the results will be much better.
So, we know from a lot of such things.
And we've been talking about it in real life, and we've been talking about it together.
Probably because of the same interest in all of what I just sent to you.
But what were your concerns from the beginning of the AI, which were you triggered?
What has you taken into it?
She was very much with Machine Learning, Deep Learning, Neuronalnetzen and so on in contact with her.
She had to use as a tool or as a component to solve, to solve certain problems.
So basically, the Ansätze in Richtung Intelligenz or at least in Richtung of a regular basis, learned or trained behavior, were not surprising.
I have already worked with Cogniti, I think they were now bought it, but it was a start-up from the Netherlands, they were also very weit in the direction of conversational training models.
So, there was a lot of points that had already been.
The rasante Achterbahnfahrt, which has been taken into the last 2-3 years, starting from these Chatbots and Chatmodels, so as you just described it, to today, the infiltration of all possible things on my computer.
The had me then again on a journey.
At first very skepticism, because...
where we have a moral idea.
We know all the issues in the right direction, the künstler's good and so on.
That's not explained.
And the AI learned of it.
The discussion in Hollywood, now where C-Dance 2.0 is out there.
All this stuff, what does now do at the top and the boundaries of the world, which were already in order and now didn't work.
That's the one part of it.
So I'm not saying that I'm not saying that I'm not saying that I'm saying that I'm not saying that I'm saying that I'm not saying that I'm not saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I'm saying that I Yeah, that's, that's what I would say since August, September, October last year, the Opus 4.5 release phase and Cloud Code, that's what I've done now really crazy in the shot and is kind of fascinant.
The parallel is very interesting, because...
Also ich habe den Kram ja tatsächlich auch studiert.
Also meine Abschlussarbeit hatte mit zu tun gehabt an der Uni und damals hat das ja alles nicht so gut geklappt.
Also es war schon mega, wie schnell du auf eine geile 80%-Lösung kommst und oh, aber elend, bis es weitergeht.
Deswegen sind mir diese ganzen Limits von, also abgesehen von gesellschaftlich, auch Energie und so weiter und so fort.
But my critique was always, that I knew where the Grenzen of technology were.
So, Hallucination, Prompt Injection, the whole thing.
And this through the statistics in the statistics, the linear algebra that is behind.
And so, that was always hard to get into it, intellectually from being here.
And then came the last November, the watershed moment, where you speak.
And that was really from now on the same for me the difference, where I said, yeah, Mist.
all your theoretical ideas, technic, are irrelevant.
Because now with the new tools, you're just hallucinating, but you can't avoid it, but you can't avoid it, but you can't avoid it.
I can't just describe it just my own way, because I'm too dumb, I can't do it anymore, I can't do it anymore, I can't do it anymore, I can't do it anymore, I can do it anymore, I can do it more with it.
But if I'm so...
I don't know, I'm doing so analysis, research, strategies and in the moment when I start to build it structurally, I don't know, I don't know what I'm doing, I'm doing research and I'm doing that not in one prompt or so, but I'm saying, oh, in a workshop you're going to go on the following 10 steps with a team and now I model these 10 steps and every step gets another context and they also agree with themselves.
Boom, am Ende hast du wirklich Ergebnisse, die so schrecklich gut sind.
Zum einen, ich gebe das ja gerade als Kurs und ein Kursteilnehmer war sehr, sehr kritisch, weil der arbeitet im Ingenieursumfeld, der entwirft Kühlfahrzeuge mit Kühltechnik, also gar nicht so weit von deiner Technologie.
Also Sensorik, Software drumherum, Hardware, Software, nicht schön, 98 Millionen ISO-Zertifizierungen.
beachten muss und so.
Ja, alles schön gut, aber wie kann ich Zeug generieren, das meine Techniker mir abnehmen?
Weil ein technischer Fehler und schon ist alles kaputt.
Da habe ich mir wirklich zur Aufgabe gesetzt, ein System zu bauen, das ihn überzeugt.
So ein Bumm, da war es.
Und er guckt sich das halt so zögernd an.
Ein paar Tage später sagt er, es ist schrecklich.
Es ist schrecklich, wie gut dieses System die Gegebenheiten von kleinen Softwareherstellern Now comes, small software-steller in the area of vehicles, in the small production production, with the Gegebenheiten of all the ISO-Norms, which they have to fill in the area of the KFZ-Bow and so on software certification, what they all over the supply chain know.
Also you see a niche, niche, niche, and they say, I can't do this.
So, and that's just the difference, where we're today.
You can build this system so, so that someone can use it in this niche, to use it, to use it with his technical skills.
Or a lot of examples, a dumbest example I make is, you have 20 interviews, you want to create a PRD, regardless of we need PRDs or not, that's not the debate.
And now you can always say, oh, Marcus, this PhD, the you generate in three minutes, the is not 100%.
And because I don't generate the PhD, but also over hundreds of phases and so on, the is just pretty good.
And now the comparison, which I always use, to calculate this technology, where she says, yeah, yeah, how possible could the person better generate.
So, jetzt die Frage in einer normalen Organisation, das, was ich dann bei Konferenzen die Frage stelle.
So, ich gebe jemandem von euch diese 20 Interviews, wann darf ich mit euch darüber reden?
Und die Schätzungen gehen dann sehr auseinander zwischen in zwei Wochen und drei Monaten.
Und jetzt meine Lebenserfahrung, wie sieht dieser erste PRD aus, über den ich mit jemandem diskutiere?
Das Problem ist, der ist auch nicht besser als der, den ich jetzt generiert bekomme.
So, und jetzt will ich den PRD gar nicht so benutzen, dass ich sage, The product manager is doof or the technology is so good.
I just want to say, how cool is that?
Because we can today already have something to discuss.
We can today already have a question about the topic and talk about it.
And if that model is even like, how doof my generated PRD is.
That's right.
for you gave a shot or kicked out?
Because you said, you said, you're a little defensive, but now, the Dammes are not going to be.
What are the observations you made, that the Dammes break and you said, you must understand everything, what's happening?
I can just say, that's exactly the effect of the acceleration of something to produce and then also in a so high Qualitätsgrad, dass es benutzbar ist und dann die wichtigen Dinge darauf zu führen, also die Gespräche, die sind in unserer Annahme wirklich richtig.
Lass uns doch nochmal gucken.
Also die Erkenntnisse, das Dokument als Basis dafür zu benutzen, dann wirklich zu überlegen, wie muss der nächste Schritt denn aussehen und wie können wir denn genau überprüfen, ob das, was da jetzt drin steht, 85 oder 95 oder 75 Prozent wahr ist.
Also das ist genau der Kernpunkt und ich glaube, der hat mich auch so ein bisschen ...
I've been working with myself, myself with personal assistants to do this.
And not in the sense of Cloud Coworks, but really something that I know more about myself, than it might be good.
Also, I know that I know from my own strengths, um that to add through certain roles and abilities.
Because that was my first approach.
I've noticed how the amount of stuff I have to do, with which I'm going to do, just more.
Unweagerly.
Then you get a PDF, and you know, that's now generated from an AI.
But you should read it anyway.
Because it's actually good.
Then comes another PDF, then comes a email, and you notice that the text says, if he has written it all, it's fine.
It's fine, it's fine, it's fine.
It's a normal effect, that then entsteht.
And then I thought, okay, I might not be able to do it anymore, but I'm only going to be able to...
Material zu sichten und sauber abzuheften.
Und arbeite gar nicht mehr.
Und dann habe ich damit angefangen und habe überlegt, okay, wie kann ich denn jetzt AI nutzen?
Und da kam dann Cloud Code gerade recht auf der Kommandozeile mit einem direkten Zugriff auf mein Dateisystem.
Das war Schlüsselpunkt.
Also quasi Schreib- und Leserechten auf meinem Dateisystem.
Und ich habe angefangen, habe eben beides gebracht, wie ich gerne meine Struktur hätte und habe gesagt, du kümmerst dich darum.
I don't know what to do.
I put it in the Inbox.
You look at what a project it's talking about.
You look at where it's going.
And that was the beginning.
And that's been a good job.
So, and there I would like to go ahead.
Because now there are two things.
That's not a person.
That's not a person.
That you can talk with the thing.
In Anführungsstrichen.
If you want to talk with him about Whisperflow or something.
But that's also a thing.
If you talk with him, you talk with him.
the Benefits, that you can give more context to understand, but you have to realize that you can influence this so.
And secondly, you have to trust that happening happening.
And as you can see, you have to trust that happening happening.
And as you can see, this is a psychological component.
There is a famous Shapiro Scale, where it's level 2 and level 3 and level 4 and 5 and level 5 is Dark Factory.
We can talk later on.
Aber der Übergang von Level 2 zu Level 3 ist, ich lese den Code nicht mehr, zeilenweise.
Ich gucke nicht mehr zeilenweise in den Code.
Und das ist halt eine Menge Trust und das ist eine Menge Engineering und Infrastruktur im Hintergrund, die einen da hinbringt.
Und der Infrastruktur muss man dann vertrauen, dass der Code schon okay sein wird auf Line-Level, dass man ihn sozusagen auf einer anderen Ebene abnimmt.
Und das ist vornehmlich diese...
this game between infrastructure and then trust building.
And that's not with, you know, from me, you're always doing this, you machine, you're doing well.
That's really, you're almost like a human relationship, what you can do to do with what you can do and learn, on which level and with which context it's good and where it's also totally schief.
And that's happening.
I don't know if that's also so.
But I have days where I sit here, where I go and don't know how it looks like it.
And I don't know how it looks like it.
And I don't know how it looks like it.
And I sit here and I'm going to say, you Schweine, yesterday you could do that.
Because the last session was in something where I'm super competent and can do that.
What have you done?
In the moment, where you said, here, file system, take it, so I would like it.
And to trust that that's already organized.
Also, have you guard rails on the ground or is there something happened, where you said, now it's too far.
Also, how did you trust to win this beginning system?
That was actually finished.
Yeah, we have from the beginning, we have said, the first aspect, the important part.
I have the whole thing in my head.
The is in my head a man with whom I talk to.
And not because I have something to do with it or something to do with it or something to do with it or something to do with it.
Also, the time is not beforegesehen.
Yeah, exactly.
But because that's in the right mindset, just like that, I don't have any emotions in the end.
Because I just say, pass on.
I would like to ask him, I would like to ask him, and thank you.
And that helps me, that is my experience, a good system to build.
I said to him, I don't want to write anything, if you have to write something, or think that you should do it, ask me immediately.
Also, bitte frag mich sofort, was du tun sollst.
Ich will auf gar keinen Fall, dass du irgendwelche Ordner löscht, irgendwas überschreibst oder sonst irgendwas.
Sondern eigentlich müsste das System funktionieren, ohne dass du überhaupt irgendwas entfernen, löschen oder irgendwas anderes machen musst.
Und das war mit einer der ersten Regeln.
Die hat er auch sehr gut berücksichtigt.
Und der zweite Punkt war, dass ich ihm auch von vornherein gesagt habe, wir lernen jetzt zusammen.
Also ich gehe nicht davon aus, dass es alles von Anfang an klappt.
Und das hat er mit in seinen Prompt übernommen.
Also wenn man sich mit Cloud auskennt, dann weiß man, dass man in dem Ordnung, in dem man arbeitet, hat man eine sogenannte Markdown-Datei, eine Datei in einem bestimmten Format.
Die heißt dann auch Cloud.
Und die AI öffnet diese Datei immer zuerst.
Das ist die Datei, die er immer im Kontext hat.
Und da müssen die drei oder vier Regeln rein.
die ich mit dem System definieren möchte.
Und das war die Grundlage dafür.
Und jetzt ist es so, wenn ich diesen Moment erreiche, dass ich emotional werde, warum?
Weil ich habe das auch, du hast vollkommen recht, ich habe auch einen Tag da, dann ist das ein Trigger für mich, ihn zu fragen, wie wir weiter vorgehen sollen.
Also das ist mein Trick oder mein Hack.
Wenn ich merke, ich werde sauer auf die AI, was irgendwie schon ganz schön schizophren ist oder ziemlich seltsam, dann frage ich ihn, was wir tun sollen.
Weil ich weiß nicht weiter.
Und ich sage ihm das auch, ich weiß jetzt nicht weiter.
Wir haben doch irgendwie gesagt, wir wollten das und jetzt machst du das.
Was können wir tun, damit es beim nächsten Mal irgendwie besser wird?
Und er ist erstaunlich gut darin, mit der Reaktion umzugehen und dann auch wirklich Vorschläge zu machen, wie man das verbessern kann.
Also es ist wirklich ein bisschen gespenstisch.
But that's the three or four points, which I have a very few points, which I have a very few points, but a clear rule in the Cloud MD, very high position, so that you read it as a first.
And then this one, if he does something like that, although I'm the one that I'm thinking, I'm actually pretty clear, I'm asking him to ask him, how we will go ahead.
That's the automatic reaction.
And that works almost too well.
And then there is a system that is already done over the time, which is now already very high, at least for what I want to do.
I have to focus a little on, because I don't want to 100% of this.
One more important rule was, I will not work with the system.
Also ich werde daran, ich werde das benutzen, ich werde die To-dos benutzen, ich werde Dateien da reinpacken, ich werde Dateien erzeugen, alles Mögliche und es kann sein, dass du nicht anwesend bist.
Also haben wir gesagt, wir bauen einen Skill, wenn ich dich wieder rufe, dass du mal eben überall guckst und mal eben deinen Index auf Stand bringst, dein Wissen über das, wo wir stehen, auf Stand bringst.
Ich werde dich auch darauf hinweisen, was ich gemacht habe, sodass ich dich wieder in einen Zustand bringe, in dem du wieder genau weißt, how our work looks like, what we're all together.
Then we can talk about the system that you built.
So, Entwicklung is a topic of experience.
Cloud Code comes on the world.
We get the great models in November.
And that triggers the first movements.
And then the idea of a system to build.
What was the idea of the system?
Wie weit hast du es bis jetzt getrieben mit dem System?
I just wanted to make the AI, so how she's also with Code, when I say, write a program, then she also has an Ordnerstructure.
They are on their best practice or on their training model.
And so I thought, so I made my Data Management with Cloud.
And that worked very well.
It was much faster than I thought.
And then came relatively quickly also that To-Do Management.
And then we realized that To-dos liegen bleiben.
Und daran bin größtenteils ich schuld.
Ich habe immer zeitliche Verpflichtungen hier und da und dann kommt es einfach vor, dass To-dos liegen bleiben.
Und dann habe ich gemerkt, da steht immer Reibung zwischen uns, weil das versteht er gar nicht.
Wenn ich gesagt habe, heute muss ich das abarbeiten, dass ich das dann heute nicht abarbeite.
Das hat ihn total verwirrt.
Also hat er dann gesagt, ja, du musst jetzt das noch machen.
Und dann gucke ich auf die Uhr ja eine Minute vor zwölf.
Nee, das war ich heute nicht mehr.
Und dann haben wir...
I started to describe why I do this.
And then I said, there is a world where I am going and there are things that happen to me, that to do.
And I have also the and the stärker and the and the schwächer.
I said to him, I'm super good at it, something new to start working in the topic, until I totally look at it.
But I'm not bad at it, if it's normal.
continuously disipliniert immer wieder dasselbe ablaufen zu lassen.
Und dann hat er angefangen und gesagt, das notiere ich mir jetzt.
Das war noch bevor die Memory-Funktion gepusht wurde, jetzt vor ein paar Wochen oder Tagen.
Und er hat halt eine Oliver.md angelegt.
Und da steht halt drin, worin ich gut bin und worin ich schlecht bin.
Und seitdem kriege ich halt auch immer schön sauber Pushbacks von ihm.
Wenn er merkt, pass mal auf, jetzt hast du vier Tage lang these things always been taken away from the beginning.
Is this still normal?
Or are you just in the procrastination mode and I'm not going to get into it?
So that's pretty exciting.
What's so exciting about this is this dimension, which is behind the end, which you could first hold for absurd.
The effect has it, when I say to the people in my course, read with the thing and ask them to whisper flow or anything, that they just talk about it, because that's so a natural interaction.
And while it's on the one side absurd, it's on the other side clear, why it works.
Because that's all of the communication is trained and on code from me.
But by the systems you describe, which you would like to do, it's just about this menschliche behavior.
That it's very good.
AIs can always look like and do as if.
Also, also, Kunst and Text, the generated from their own strategy text, but a text that looks like a strategy text.
And, of course, it works.
So, a super trick is to say, first of all, consistent in the model to stay and say, oh, there is something about what I'm talking about.
Also, I just explain what I want.
Fertig.
And the crazy thing is that...
non-deterministisch, eine Übersetzung dessen, was ich will, irgendwie stattfindet.
Und das ist eine Blackbox, wir verstehen es nicht genau, außer wir können genug lineare Algebra und so.
Aber das Modell zu haben, dass ich mit dem Ding einfach rede und sage, was es will, ist ein Riesenschritt, wie man die Dinger bedienen kann.
Und jetzt kommt es.
Was oftmals überhaupt nicht funktioniert, ist das Gegenteil.
Also sehr, sehr funktional mit den Dingern umzugehen.
Also Beispiel, letztens habe ich Open Claw.
Ich habe mir dann von Claude Code helfen lassen, OpenClaw aufzusetzen, weil das für mich ein viel geileres Interface war, als auch nur anzufangen, in den Config-Files rumzufummeln.
In dem Moment, wo man das System bricht und mal auf so einer Verhaltensebene spricht und dann wieder sehr technisch wird, geht ganz viel wieder kaputt.
Also mit einer Worte, OpenClaw manuell konfigurieren geht, konfigurieren lassen überreden geht auch, der Mix-Modus ist immer verheerend.
And that's my experience.
You have to be able to get into the description of the thing, with which you talk about, relatively consistent and consistent, so that you can build a long-term relationship.
And what many people are wrong, in Anführungsstrichen, is that it works not.
Now they go low level and start to make it on a file level, everything directly.
and wonder if there are some weird effects that come out.
Why?
Because the system is not thought about it.
You have to build it differently, that you want to interact with the files.
That can be done with the files.
But I think it's always good to stay in one mode with the system.
And that's, I think, a difficult problem for the programmer.
I don't care about it anymore.
So wie heute ein Programmierer sich auch nicht mehr für den generierten Binärcode interessiert, was früher halt nicht selbstverständlich war.
Aber wenn du heute anfangen würdest, einen Binärcode rumzufummeln, bist du verloren.
Das geht einfach nicht.
Und ich glaube, diese Abstraktionsschicht zu akzeptieren, ist ein Riesenschritt.
Verrückterweise.
Absolut, da stimme ich dir total zu.
Das ist auch meine Erfahrung.
Und das hat mich so daran gebracht, dass ich sehr stark damit experimentiere, wann ich Kontext...
and I need to use and what I want to use context.
Ich fange nie an einem Softwareprojekt, wenn ich nicht sofort mit meinen schon vorgefertigten Assistenten starte.
Also ich habe quasi eine vorgefertigte Cloud-ND, eine vorgefertigte How-We-Work, eine vorgefertigte Wie funktioniert das?
Wie gehen wir mit Git um?
Wie machen wir das?
Wie machen wir das?
Also das ist alles schon da.
Und wenn ich quasi ein Repository starte, dann lade ich als allererstes mal die Datei rein, starte Cloud und sage, das machen wir jetzt.
Und manchmal will ich das genau nicht.
Manchmal will ich, und das ist vielleicht so ein bisschen mehr der Vibe-Coding-Modus, will ich einfach mal ausprobieren, was denn passiert.
Also ich bleibe sehr vage, ich versuche gar nicht so genau zu beschreiben, ich versuche gar nicht, in so eine Art spec-driven zu gehen, sondern lasse mich fast schon so ein bisschen darauf ein, womit er zurückkommt oder sie zurückkommt.
Und mit diesem Kontext zu spielen, halte ich für extrem wichtig.
Also zu sagen, ich will gar nicht, dass du weißt, was wir alle schon geteilt haben, darum bin ich auch gar kein so großer Fan von dieser Memory-Funktion, die Claude mittlerweile ja fest eingebaut hat, weil der hat ja immer Kontext.
Und manchmal willst du das gar nicht.
Manchmal willst du ja ein Pushback oder eine Challenge oder vielleicht auf irgendwas völlig anderes kommen.
Und darum spiele ich sehr viel mit diesem, vergess mal deinen Kontext jetzt und challenge mich mal in einer völlig anderen Rolle oder so.
Yeah, with memory, I like Kurzzeit-Gedächtnis.
Also, I find it great, that they can in the meantime Sessions overbrücken with what I want.
But I find also Langzeit-Gedächtnis often relatively schädlich.
What I want to say, what you just said, my observation is that the best Interaction is always on the output to look and say, the guy likes me, I would like him so.
the instructions description, not so much to make, but really to say, yeah, it's pretty good, but if you look at the 3rd, or this section, or whatever the report is, or the file that comes out, in the fifth section, in the fifth section, in the fifth section, there is a lot of this and that.
And then completely to realize how it's going to go there.
And that's the very interesting thing, that the suggestions, how it's going to get it.
often so to the point and also to the point and also to the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the point of the Und wo kann ich das auf der anderen Seite überhaupt nicht voraussetzen?
Manchmal ist das so überraschend.
Eine der größten Überraschungen, die ich immer wieder erlebe, ist, wie wenig Claude Code über Claude Code weiß.
Und dass Claude Code mir jeden Tag 20 Mal die Claude MD vollmüllen würde.
Und ich muss jedes Mal da stehen und sagen, wir wollen doch nicht so viel in die Claude MD schreiben.
Ja, aber es ist so bequem.
Ja, das mag für dich bequem sein, aber wir machen das nicht.
Ach so, stimmt ja.
Genau, aber ich glaube, das...
This outcome driven being and the outcome to describe the outcome, not so stark the constructions we can describe and this interaction with the LLM.
Where can they go?
I think that is totally the key of this trust in the world.
Yes.
And on the way, that can you do it with the open questions.
in einem halben Jahr noch niemals darüber nachgedacht hätte, sowas überhaupt eine AI oder ein LLM zu fragen, dass du sagst, du hast mir jetzt hier einen Spec geschrieben mit ein paar Use Cases und wie willst du denn jetzt beweisen, dass das für den Nutzer wirklich wertvoll ist?
Und das versetzt ihn, was dann natürlich passiert ist, dass er relativ tief in sein Trainingsmodel geht und wirklich nachgucken muss, welchen Wahrscheinlichkeitspfad finde ich denn jetzt, um so eine Annahme nachweisen zu können.
Und das ist schon, also das ist wirklich sehr interessant, mit welcher Güteklasse der oder sie da teilweise zurückkommt und versucht, diese offene Frage zu beantworten.
Ich habe mir lange Gedanken gemacht, why someone like Boris Czerny, the Papa of Cloud Code, why the always predates that he personally, that must be sure not, but he is always long in Plan Mode in relation to others.
And I think I have a explanation.
What I have found is, if you have a relatively generic LLM, give me a review to this and this, then comes a relatively generic, doof review.
Let's have to do a bootcamp review.
And that review was relatively good.
I knew before, there were no mistakes in the end.
I was going to go on this review.
And then I said, look at the point where this and that is described.
And then he looked at the point where he said, the point is exactly wrong.
And then I said, wait, is that now biased because I'm going to go on?
Then he explained to me, no, I have now first that Fachwissen has been done.
And that's totally important to understand.
LLMs are not so that they always know everything, but you have to throw them in the way.
And say, look at them, look at them.
And I think that's both in...
So if you 20 Agents hinterher los send in the same session from me out, every Agent zapes a different speciality.
And if we're in Plan-Mode, what we're done is also, that we say...
Guck dir mal dieses API-Wissen an, guck dir mal jenes Strategie-Wissen an, guck dir mal das an.
Wir zwingen die LLM, die verschiedenen Spezialwissen reinzuziehen, dann qualifiziert darauf zu gucken, als wir das in einem Prompt jemals einfach machen könnten.
Ich glaube, das ist total wichtig zu verstehen als mentales Modell, was auch den großen Unterschied ausmacht zwischen dem Rumgechatte und halt konstruktiv diese Verfahren Stück für Stück aufzubauen, den Kontext gezielt reinzujagen und so.
The LLM must be able to get on the way you look at it.
Yes, that's so important.
My biggest fear is the loss of AI.
The storyline is very simple.
AI makes more of what we did before.
And we learn to do it in the way, that we can't.
We can't do it in the software.
We write our software and you learn it.
The principles of software.
You don't understand it anymore.
There is a problem.
I can't even say anything.
I can't even sell it anymore.
I learned that probably 100 years ago all people can't.
But no matter.
There is a problem.
I have a new assistant.
That's a learning assistant.
He should be able to I can give up to the curriculum, in which I can do with self-paste things myself.
Also from music theory, it can be all possible.
It's me egal.
And I build it right.
And then I have to tell you, if I go to a research, a research is, researche please, how the person learns.
And then comes a relatively generic...
thing back.
It's already good, it's good to read, especially with Opus or Sonnet.
I find the language that is very impressive, the models, it's good to read, you can read it, I understand, but it's on the top.
And then I said, well, good and good, but I want to go more.
Please take that document auseinander and for every theatical part, the you find, also Feynman, Borowski, everything, Mach bitte nochmal ein Deep Research speziell zu diesem Thema.
Und der braucht am Ende 14, 15 Minuten.
Also erst lässt er dieses generische Ding laufen.
Das packt er dann im D-File auf meine Festplatte.
Und dann startet er bis zu 10 Agenten und nimmt diese Einzelthemen da raus und geht damit auf Deep.
Und das, was du da zurückbekommst, ist so unglaublich.
Titles, könnte ich als Buch lesen.
Ohne Mist.
Also das ist so krass gut.
Und er macht natürlich den Research im Web, aber er hat natürlich auch viel im Trainingsmodell drin.
Also das ergänzt sich dann ganz gut.
Aber dieser Unterschied zwischen mach mal einen generellen Überblick oder mach mal genau, beschreib mal, was Feynman geschrieben hat, ist so ein himmelweiter Unterschied.
Und das ergänzt sich so gut.
Darum, genau wie du sagst, du musst immer wissen, auf welcher Ebene On which layer operiert der gerade?
Operiert der gerade auf der, hey, ich bin Claude und ich weiß eigentlich alles auf der Welt?
Oder ist der gerade schon irgendwo in einem Keller und ist da quasi in so einem Expertenmodus drin?
Das ist so wichtig.
Leute sind immer wichtig, süchtig nach Use Cases.
Was kann denn dein System umso alles?
Gib mal ein paar Beispiele, dass man das irgendwie ein bisschen eingrenzen kann, was dein System so macht und was es nicht macht.
Also der Assistent, der hat mittlerweile I have several roles from me.
We have a how we work.md.
There is a way we work together.
And I'm going to differentiate between Agenten and Rollen.
The role can he take every time, he can he change himself.
And Agenten are the way I'm going to go to the point.
So I'm going to say, I'm going to go to the point and make it.
And the role is the...
He's my assistant, he's my system on the way to hold him.
He's going to see if we have a new threat, if I didn't bring anything to the end, what I wanted to do.
He's going to hold him.
I have a David Allen getting things done.
That's it, I can do things in the future delegating and they come back to me.
He reminds me...
He also tells me, if he gets the impression that I ignore something too much, that's he.
We call it Nudging, then he can just say something.
And that's the basic foundation.
Then he's a strategist consultant.
I gave him a question.
That's a knowledge file.
I was in the system.
I asked him to think about this whole idea of how organizations are built up.
This management phase of the 50s, 60s.
I made several researches and said, you are a strategic consultant.
You know a bit the history, you know, where the companies are developing, you know what the right thing is for us all.
And if we go in a strategic direction, then I would like to give you a pushback or a info to something.
Then he is the Wissenskurator.
There is a relatively wide range of Wissens, a Wissensdatenbank, could be said.
That are only just the data.
An die darf er nur drangehen, wenn er sie wirklich braucht, weil es macht keinen Sinn, die in so einen Kontext zu laden.
Das ist zu viel.
Aber wenn ich ihm sage, lass uns mal genau bitte in das Thema gucken, dann darf er halt auf diese Wissensebene runtergehen.
Und das soll er bitte auch, also da hat er die Aufgabe, darüber einen Index zu führen.
Also wirklich immer wieder, es gibt auch einen Index-Skill, den ich laufen lassen kann, weil manchmal schiebe ich da halt auch Dateien rein, während er nicht läuft, also wo er das gar nicht mitbekommt.
And then he's Tool Builder.
So that's when I say, a tool is, I gave him the corporate identity requirements of the company.
And I gave him a PowerPoint, if I want to have a PowerPoint, then are those your requirements.
The color, the writing and so on.
And he built a Python skill.
I can't ask him to do that.
And he can say, I can't do that.
And he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he knows that he And he has basically a tool for it.
And that are the five or four, the four, the one you can take every time, depending on the demand, the we have.
And then there are three roles, which are a little bit about it.
That are the right agents.
And that is the, take a different view of mine.
That is a agent.
Also I say to him then, please...
I have to give you my perspective, I have to give you a perspective, what I think, what I want to do.
Make a pushback.
Just take a different perspective.
It's not a matter of you finding it, it's not a matter of you finding it, it's not a matter of you finding it or not.
I want to have a second opinion to this topic.
Then I have a CEO skill.
Also I have to research them.
How they behave?
What is the task of a CEO?
How do they look at their job profile?
What is their responsibility?
How do they have to behave?
How do they have to behave?
How do they have to behave?
And that's how he has to behave.
And I said, if I want to say that, then you're my CEO.
And you're going to challenge me.
I'm going to get a presentation.
I want to tell you about something.
I want to make a pushback, whether I come through or not.
And that's very good.
Because I don't have a business like a business, if I have ideas.
And he killed me just like this.
He said, Oliver, what do you do?
I need to pay.
And then I know, I'm not so close to it, as I thought I would be.
That's the most important agents I have.
With the part that we take care of, the Review passiert so.
Immer mal wieder zu Reviewen ist das immer noch so, wie wir, also wenn wir feststellen, dass wir nicht so richtig weiterkommen oder wenn ich merke, ich bekomme jetzt nicht das Pushback, den ich vielleicht erwarte oder so, dann iteriert man halt auch nochmal darüber.
Aber das sind die wichtigsten Agents.
Wie sieht denn deine Interaktion aus?
Also ist das, also offensichtlich ist es ja irgendwie wichtig, dass sich das Ding irgendwie Dinge merkt, Zustände merkt.
where you are currently working or not working.
And the other thing, it seems to be a kind of scheduling to be.
There are many systems that are just skills-based.
I'm calling you skills and then happens something.
That seems to be a bit of activity to be.
And then as a third element, at least at the end, there's also Push.
That's from itself on and says, Oliver, you've got something to do.
Or we know the devil, what he does.
Can you describe how the interaction is?
And also the channels, with which you interact with.
Why do I ask that?
My experience is that you can achieve something with Scheduling low level with Conjobs or so.
All that you can do, if you do everything in Claude.
I don't know exactly.
It's still a bit shaky.
Also, it's not a bit shaky, it's not a bit shaky.
Also, if it's not a direct self-made job.
And Memory is the same.
Also, a lot of things you describe, is in my experience still bleeding edge and not always so, that it bulletproof is.
Maybe you can a bit of a bit about it.
But, why?
Because probably many of them are trying to try it.
And many of them are trying to First of all, the memory is happening.
The thing happened in the session, but it was not written.
And then the next one flasht man on it and then it's gone.
That the schedule didn't come.
The famous morning brief, which they all have written.
Today it doesn't come.
What's happening now?
But there it is.
But he was on.
Where is it?
MCP-Server not reached.
Schedule had not started.
Also, how...
What is your learning process and your learning process with this bleeding edge, which exists?
That's a little bit of a problem, if we're just in the research mode or in the work mode.
Also the work mode is really trivial and relatively rare and possibly unanfällig.
I have a Hello-Skill and I have a Goodbye-Skill and when I make the commanding on...
Da fühle ich mich auch wohl.
Also da habe ich keine Berührungsängste und gebe ein Slash-Hello ein.
Und er führt so ein Startprogramm durch.
Er guckt in die To-Do.
Er guckt, was ist von letzter Woche liegen geblieben.
Er schiebt alle To-Do's von letzter Woche in die aktuelle Woche.
Er erinnert mich, pass mal auf, da hast du ein Meeting, da hast du ein Meeting.
Also das, was ich so als Reminder mal angelegt habe, die macht er dann auch.
That's the basis for the first time.
But then you have to go to the next level.
You open Terminal and start the Claude and then you say hello.
In other words, the session begins new.
He knows not much, except what you know.
Your mental state is, you can only take care of what he persisted.
He has no existing session-session-wissed and now you say hello and so you initiate the day or the day or the day or whatever you're doing.
Genau.
Also er hat...
vergisst er oder löscht er auch wieder nach einer gewissen Zeit.
Das Change-Doc heißt bei uns Sessions.md.
Und da sage ich ihm, und da habe ich immer noch nicht von losgelassen, da sage ich ihm explizit, merkt ihr das in der Session?
Beim nächsten Mal kommen wir vielleicht wieder darauf zurück.
Und die durchforstet er auch in dem Hello-Skill und guckt kurz, alles klar, letztes Mal haben wir das und das gemacht.
Today is, what we know, Friday, the so-and-so-fielte.
You have two To-dos, which you want to do today.
Today is the end of the day.
You have a Termin.
So, there starts with.
And also, I don't need to do a brief or something, but it's a very short time.
And he looks a little bit on, is it a week-end, is it a week-end?
Because at the week-end we're doing the most strategy work.
And in the week-end is really the mode of work.
And that's the Hello.
Hello-Skill.
Den kann ich aber jederzeit laufen lassen, wenn ich der Meinung bin, wir haben jetzt ganz tief an einem Thema gearbeitet, jetzt wollen wir da wieder raus, starte ich wieder Hello und der fängt quasi wieder an und sagt, okay, das haben wir gemacht, das machen wir als nächstes.
Und dann normalerweise ist es so, dann schicke ich ihn in irgendein Projekt.
Also wir beschreiben uns, beziehungsweise mittlerweile kannst du auch auf der Kommandozeile die Spracheingabe nutzen.
You can say, we work on this project, product portfolio or so in the way, look what we did last.
Because every single project has also a session MD.
There are also the things that we did in the session.
And make yourself smart and then we start.
And then he comes back with, I have seen, we have done this and that.
As soon as you want to do this, what do you have done?
And then I say, I have the input.
I have the file in our input-order, I have to hold them, I have to read them through and then let us see what we have to change our concept.
On the basis of the new experiences.
I don't know, I have any sales sales or I have anything, what has to do with this project.
Please read it through and let us see what we have to do with the next steps.
And that's worked very well.
He is super self-conscious and comes then with...
weil er eben so ein bisschen Kontext im Bauch hat über, was ist denn das für eine Firma, für die wir arbeiten, was ist denn der Oliver für ein Typ, was haben wir für Strukturen, kommt er mit relativ guten Resultaten auch relativ schnell zurück.
Und dann gibt es den Goodbye-Skill und der Goodbye-Skill ist im Prinzip nichts anderes als, jetzt nimm mal deinen ganzen Kontext und dokumentier den mal weg.
It can be, that it's the global session MD, it can be, that it's a project session MD, it can be, that we have a knowledge index, because two new knowledge data, two new researches, come to the end.
I have a name for a date.
For example, K at the beginning with a minus behind, is a knowledge data, is a research or something.
Then a TPL is a template.
Also es gibt so eine Namenskonvention für Dateien.
Dadurch muss ich ihn nicht immer jeder Datei lesen lassen.
Das ist ziemlich wichtig, weil da verbrennt er halt Tokens ohne Ende.
Darum habe ich auch diesen Index, dass ich mir halt direkt sagen kann, guck mal bitte, bevor du jetzt hier ein Crawl durch mein ganzes Verzeichnis machst, schau bitte in den Index, wo wir das und das haben.
Und er holt sich dann quasi auch direkt die Datei.
There I have at the beginning quite a lot of times with a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit of a bit The one who I'm planning to plan is, I don't like it because he tries to get quite fast in a question-of-answer situation and I will talk about it.
And he will talk about it.
Yeah, yeah, yeah.
A little bit more.
Yeah, yeah.
Yeah, yeah.
Yeah, yeah.
Yeah, yeah.
Yeah, and this kind of compound effect, also that means that he always has a session with a new certificate, new where he writes, if not only in short words, the one is super, super important.
And I notice how fast...
he then will lose the Faden when I do not do that.
So it happens when I make the phone to the phone, make the phone to the phone, make something else, make the phone to the phone, I forgot the goodbye skill.
Then the context is still there when I do it again, but the time is still there, until he has the whole thing again.
And if you call hello on, Is he really in a different situation or is he just talking about it?
He has just a lot of information in his context.
Or if you go into a project, what happens physically with him?
That has to do with the structure, which we started with at the beginning, which I have set up.
I have used the sogenannte Para-Ordnersystem from Tiago Forte.
That is a relatively simple system.
on the upper level are actually only four Ordner and everyone has a semantique.
There are projects in the Ordner, there are all projects in the area, which are now limited to the time.
They will be for a certain time there and then probably somewhere else.
Then there are areas, which are things that I always have to do.
That is my job, there are my Dauer-Aufgaben, there are my Dauer-Sachen.
Then there are resources, that is the whole Wissensdatenbank, everything.
from templates, over cdci, over regels, over processes, everything else, and then there is the archive, which he can use, to make things away, if we don't need it.
And this semantics alone is so helpful, because he knows exactly, if I say, pass up, we're talking about this project of Swiss Learn, then he looks at the project and looks at this project, or he looks at this project.
And this is very important, because Das erzeugt im Prinzip vier verschiedene Ebenen.
Die oberste Ebene, das ist die Cloud MD.
Die hat er immer, die weiß er immer, die ist immer in seinem Kontext.
Da stehen die wichtigsten Regeln und Verhaltensmaßnahmen drin.
Darunter kommt How We Work, Sessions MD, Change Log, also das, was ich vielleicht noch wissen muss, was ist vorher passiert, was wollen wir als nächstes machen und so und To Do.
Darunter kommt so was wie Projekte, dann sind wir schon relativ tief in einem Thema drin.
Und darunter kommt diese ganze Wissensbasis.
Wenn ich halt sage, guck mal bitte nach, wie bei uns die Reiserichtlinie ist.
Ich muss, keine Ahnung, nächste Woche nach Berlin oder sowas, sag mir mal eben, wie bei uns Reisen beantragt werden.
Ist jetzt ein bisschen erfunden, aber sowas könnte der halt auch.
Das heißt, durch ...
Dadurch, dass Para relativ prominent ist, hast du ihm schon so ein Regalsystem gebaut, wo es immer weiß, in welches Regal es gerade überhaupt greifen soll.
Und wo es sich in dem Regal dann noch orientieren muss.
Genau.
Und ich glaube, das ist wieder so ein relativ schlauer Übergabepunkt, wo du halt sozusagen eine Sicherheit anbietest von die Navigation kann nicht vollkommen falsch gehen in den Kontexten.
Weil diese vier Überbereiche And that's very good from the LLM.
If you speak about a project or resources.
Yes, on the way.
Especially through this Semantic, the there is, it's very easy to understand and for him very easy to hold.
And the helps me just to, that I don't have to do it in the navigation or the finding or the able to do it.
That's super, super important.
Let's just talk about the interfaces you use.
OpenClaw has this seemingly bestiehful Wunsch and has been recognized and has been used to it.
Where I am, I want to interact with my agent.
The Americans say, go touch grass.
Also, go with your Hund and work there.
So ein Cloud Code baut das ja jetzt auch Stück für Stück nach.
Es gibt verwirrend viele Interfaces.
Du kannst über Channels, Telegram mit ihm reden.
Meine Erfahrung, wieder nicht so stabil, beziehungsweise nur stabil in einer Session und mal gucken und so.
Dann gibt es Dispatch in Cloud Cowork, wo du Remote zugreifen kannst.
In Cloud Code selbst gibt es Remote, was aus meiner Sicht stabil ist.
Also ich zum Beispiel habe...
I have every time two sessions on my phone, so that I can always use the sessions that I can always use the same thing that I found to be able to do cloud code.
How do you interact with your system?
That's a bit dependent on the company and on the company's computer use the stabilest variant, the terminal.
Und ich bin im Terminal.
Und dann nutze ich maximal die Stimme.
Also das heißt, entweder tippe ich so oder so.
Wenn ich in einem Meeting bin und ich will kurz was fragen, dann tippe ich natürlich.
Oder ich nutze Stimmeeingabe, was ja mittlerweile auch auf der Kommandozeile funktioniert.
Und das reicht mir vollkommen aus, auch für das, was ich da an Informationen verarbeiten kann, was von Ihnen zurückkommt.
Für diesen Firmenanwendungsfall.
Ich habe damit experimentiert, den in Slack einzubinden.
Also in ein Chat-Tool und da auch einen Webhook einzubauen, dass der einen Trigger bekommt, wenn ich mit ihm reden will.
Das hat auch sehr gut funktioniert.
Allerdings ist das eine andere Art der Kommunikation und ein Stück weit auch langsamer, als wenn ich mir halt im Terminal irgendeine direkte Anweisung gebe.
Darum greife ich da eigentlich relativ selten darauf zurück.
Zu Hause, also auf meinen Privatrechnern, I have not so much an assistant, because I don't have this super professional application, that I have to do a lot of seconds, but I can use it a little bit more.
But I use it a lot of channels.
I use it like you, a remote-accountable.
I use the Loop-Incognitive-Incognitive-Incognitive-Incognitive-Incognitive-Incognitive-Incognitive-Incognitive-Incognitive-Incognitive-Incognitive-Incognitive-Incognitive-Incognitive-Incognitive-Incognitive-Incognitive-Incognitive-Incognitive-Incognitive-Incognitive-Incognitive-Incognitive-Incognitive-Incognitive-Incognitive-Incognitive-Incognitive-Incognitive-Incognitive-Incognitive-Incognitive And I had tried to do that with OpenCode.
I tried to put the whole thing in Visual Studio Code to hang out, so that you can get the management, the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of the management of OpenAI hat ja Symphony herausgebracht.
Im Prinzip ist das ja so eine Artrello Kanban-Board für Agenten.
Jetzt gerade vor einem Tag hat Claude auch a Management Agent oder Managing Agents angekündigt.
Es gibt da schon die ersten YouTube-Videos.
Das ist tatsächlich nochmal interessant.
Das ist halt eine Kanban-Anwendung, die im Prinzip auf einem Localhost läuft und in dem ich quasi...
over Kanban-Karten, certain Agenten in a process, in a flow, or even a Super-Agent, a message can and he moderates the Sub-Agent.
That must I have to see now again.
Symphony found I a little trage and a bit strange and that was for me very strongly on this coding-anwendungsfall.
Mal gucken, wie jetzt das Managing Agents von Claude wird.
Das könnte ich mir halt vorstellen, dass das nochmal ein bisschen interessanter wird.
Wenn ich jetzt Revue passieren lasse, dann muss ich irgendwann hingehen und muss mir ein Always-On-Claude irgendwo hinpacken.
Bisher ist der Claude immer dann an, wenn ich den Laptop aufmache.
Und es könnte gut sein, dass ich jetzt irgendwann, ich habe keine Lust auf OpenClaude.
I tried it too, but it's too risky.
But it could be that I could set a Mac Mini for a while to do that there's a cloud that can continue to communicate with the cloud.
I had two weeks invested in Open Claw and let it go, because it was a high maintenance for me.
For the first time, endless Token costs.
because it's also the context always so on, because it's in a index, the session holds and always the whole index with it.
So you come, even if you always make the session compact, you come to 20k, 40k tokens, with every question.
So for me it was so, that every telegram question had me to get with Sonnet 20 cents.
That was then albany and so.
If you think about it, you talk about it or not.
Anyways, the warning is also about the managed agents of Claude, they are going, so as I know, over the normal API costs.
And that is just for every normal private user, is just a lot of what is available for the Abo.
So that you should also have to pay attention to what costs are.
I have it yesterday once tried.
So the upper surface things, which you can do, are...
trivial.
I think when it gets used to be more difficult, what works with them.
I think the real break at these managed agent platforms is that it is just hundreds of startups in question, which are currently on the market, to build this infrastructure, which is given to the API costs.
The whole infrastructure is a whole.
Probably some of the people who have thought about their future is, if the big companies are doing this all with the handle.
That's a great impact.
Absolutely.
And I think that's a spiral.
We'll move on to the next.
Ja, mal gucken, wo die Reise dahin geht.
Ich habe tatsächlich, wie gesagt, ich komme dann immer wieder zurück zu den Basics.
Da bin ich auch so ein bisschen mit am schnellsten.
Ich finde auch, das Coding bei Telefonen ist natürlich fantastisch.
Also wenn du dann mit dem Telefon sprichst und im Hintergrund eigentlich mit deinem Rechner, also dieses ganze Remote-Thema.
Was mir so wichtig ist, ist, ich habe relativ wenige sequenzielle aufeinander aufbauende Aufgaben.
Mir ist fast eher das gemeinsame Konzept und die Diskussion so wichtig.
Danach, ihnen dann irgendwas abarbeiten zu lassen oder zu sagen, mach doch mal einen Plan mit fünf To-dos und dann schick mal drei Agenten da drauf, das ist trivial.
Aber bis ich erstmal da bin und weiß, was will ich denn überhaupt genau machen, das ist eigentlich das Wichtigere daran.
For everything, what not coding is, is that, I believe, another super important point.
Today morning in my course, we talked about it.
I had last so a thing as Agent gebaut.
And all of them said, cool, that it goes, but it's not my use case.
What makes a agent?
The rhodes the 500 steps autonom through.
For him.
What the people said, there have I no idea.
For me is cool in my job, Research-Phase durchröten lassen, Zwischenergebnis kriegen, nächste Phase beeinflussen, nämlich das Research-Ergebnis nochmal sozusagen drei Informationen mitgebe, jetzt sich auf die nächste Phase begeben.
So und dann war unser gemeinsames Fazit, die Idee, wie kriege ich den Human out of the loop, ist fast immer komplett langweilig.
And that you read and you hear and so.
But actually it's a completely wrongful case.
That's where you really have something like, I don't know, a break.
That's not what you do.
But in the moment, where it's about, oh, I'm going to talk about people, I'm going to talk about communication, is it not that you have to do anything that you have done a research that you don't understand, then you don't understand these 33 pages, which come out, you have to understand, without the way to get to it.
Because the next situation in the company is, You have to explain why you have final accountability for these 33 pages.
So, why do I have to do this?
Why do I make a whole life-saving?
And I have to do this.
So, what you said, is this important thing.
In real life, I have to do this.
There are almost all cases that we need to observe the steps that we need to observe and then again to some stakeholder on the world.
To the chef, to the staff, to the peer, whatever else.
And that's for me a very interesting experience, that all these automatization-warns, the behind-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the-the Wo muss ich aber reingucken?
Wo will ich es trotzdem verstehen?
Weil hier ist die Human-to-Human-Grenze, die sowieso nicht weg bleibt.
Weil die Fantasie, dass du Unternehmen leitest, dass du Dokumente generierst, die automatisch analysiert werden, dann eine Maschine vielleicht die richtige Aktion ableiten.
Das gibt es halt nicht.
Genau, und ich finde es fast wichtig, also das ist genau der richtige Punkt.
Viel produzieren, das kann jetzt jeder.
Das ist kein Alleinstellungsmerkmal mit.
Am Anfang freue ich mich natürlich, ich brauche mal 33 Seiten Research und das ist alles ganz toll.
Aber viel produzieren kann jeder.
Aber dann diese Geschwindigkeit wieder übertragen in die Organisation, in der ich arbeite, das ist...
the key to use and then with good or at least with good results in the next step to go, to go to the alignment to go, to go to the people, to go to the people, to go to the people.
That's actually the key to the point.
Not that many produce.
Also from my real life, I was thinking before AI, that you have a document where you put it and people understand it.
B, they understand it like you.
And C, perhaps even more gewager, they are even motivated or even motivated.
And all four points are first of all a lie.
I think that's one of the wonderful machines that can be understood by all the wonderful machines, which can be understood.
If you read about how management should not only be supported by AI, but also being used as glue in a company through AI.
I think what you always forget is that we always talk about sozio-technical systems and not just technic.
And that it really is not a skill from AI, to set up the decision in order to set up and to set up the decision in order to set up.
I think that's what we always have in mind.
That management, especially in the middle management, can very, very, very, very support.
But it's always the...
Human-to-human, Kommunikation, Ausgleich, Alignment.
Und ja, viel funktioniert über Infrastruktur.
Also du und ich können Gespräche haben, dann können wir die MDs in unsere Produktionsstraße legen, die hält sich dran.
Alles cool.
Aber trotzdem brauchen wir beide erstmal das Alignment, welche MDs mit welchen Aussagen, welcher Semantik denn in der Produktionsstraße enden, damit das richtige Zeug produziert wird.
Und dann produziert das wie der Teufel.
Kein Ding.
So, aber ohne dass wir diesen Input geben, passiert es halt nicht.
Genauso wie halt die AIs nicht morgens aufwachen und sich MDs generieren und Output generieren.
Das machen sie halt nicht.
Und wenn, ist es wahrscheinlich auch nicht schön.
Bis jetzt.
Also AI macht halt nur können, nicht wollen.
Das ist halt der große Unterschied.
Wollen müssen immer noch die Menschen.
Und da liegt halt immer noch der Kernpunkt jeglicher Motivation begraben.
I think it's important to be more invested in this.
That's what I've done with.
The AI has been in its own constructs, you to help.
That's a lot of confirmation bias.
Everything, what you say to the AI is at the end, is right.
If you work with something, you ask the AI, was that good, what they did?
Yeah, but that's good.
We have just a hour of work.
Of course, that's good.
And then, but...
the AI, the reflection, or even the conflict to give.
Also to say, I want to say, I want to say, I want to say, I want to say, I want to say, I want to say, I want to say, I want to say, I want to say, I want to say, I want to say, I want to say, I want to say, let's say, let's say, let's say, let's say, let's say, let's say, let's say, let's say, let's say, let's say, let's say, let's say, let's say, let's say, let's say, let's say, let's say, let's say, let's say, let's say, let's say, let's say, let's say, let's say, let's say, let's say, let's say, let's say, let's say, let's say, let's say, let's say, let's say, let's say, let's say, let's That's a important aspect that I have learned in the system and in the system in the system.
There is a incredible Bias in the system.
Everything is right, what we do.
Everything is good, what we do.
Even Claude or Opus and Sonnet, which are still not so much like ChatGPT.
Selbst da merkst du, irgendwie ist jede Idee gut und toll und gut, dass wir das geschafft haben.
Lassen Sie das machen.
Ja genau, und du brauchst ganz dringend immer wieder auch mal den, der sagt, ja Moment, seid ihr wirklich sicher, dass ihr das machen wollt?
Guck mal, da gibt es noch die und die Punkte.
Jetzt mal eine angesagt kunstvolle Überleitung versuchen.
Handbag had before a few or three years already said that 90% of his skills reduced their value to zero and the rest of 10% of his skills reduced their value to 1000% now where AI comes.
And he talks about not so much about AI but about Genies who help us.
And what he said is that the personal challenge and the business and the society is the challenge.
these 10% and the 90% to clarify.
What are they in my own context?
What are those for students, students, people who develop software, people who work like me and so on.
What are the 10% of them?
Because these are ultra cool and these are mostly also the most human.
Now, as I said, the overview.
Now you have a little bit described how you crystallize this system, how you can't get 90% from the 10% from.
You are a responsible person in a business, also in a business environment.
What does your company make so, um this knifflable job to get?
That's very difficult.
There is not a...
For everyone of us is difficult to do that, to just to keep the developers in the way they do it.
Because what they have for 15 years defined, about Code to talk, the they generate, the word generated, what is it, what I have been doing 15 years.
Where I talk about the dinner, the dinner with colleagues and where I can even go through religious wars.
How do we do the variables?
And so, haha, and the one makes it wrong, haha, but I make it right and so, boom, away.
How do you do it with this 10% and the 90%?
And what are the challenges you get?
Or what are the first steps you found, to be responsible for that?
Because that is what we are looking for.
How do we do it with responsibility?
Not because we are just people and want to work with other people.
You had a big advantage on the way we are dealing with it.
First of all, I have to say, in Testo is it also so, that we have a very strong difficulty in the Fähigkeiten of the use.
The development of the software is of course already extremely involved and it has to be there with really a lot of detail.
Then we have other areas.
We use AI or start that whole thing, but we test ourselves first of all, because the direct value is not immediately recognized.
It's not so, as if we're the AI-Götter or so.
In the software development is actually so, you say often, Cadence is super important.
where we are going to go, to see, to see, to see, to see, to see, to see, to see, to see, to see, to see, to see, to see, to see, how we are going to go now.
And that's almost a mantra for us at the moment.
It works well, well not so well, but the Rahmen and the Rue to create, that we talk about what we are doing, is a thousand times important.
I can...
That were 10x developers, they were super good, now it's 100x developers.
They run like a mess, they use a code, but it's not a thing.
Because the alignment between the developers, the architecture, the decisions we have to make together, the knowledge transfer, there must be many spaces.
That's what were the 10%.
Man wollte gerne, dass das nur die 10% waren und 90% das Programmieren.
Und jetzt dreht sich's.
Ich habe 15 Agenten am Laufen und 4 Agenten steuere ich direkt.
Und die programmieren das alles.
Und da bin ich in einer Stunde so weit, wie ich sonst in einer Woche gewesen wäre.
Aber jetzt muss ich mich unterhalten.
Jetzt muss ich zu meinem Architekten.
Jetzt muss ich zu einem anderen Team.
Jetzt muss ich...
Guck mal, ich habe eine Schnittstelle gebaut.
Kannst du damit was anfangen?
Oh ja, lass mich das mal testen.
Geht das in die richtige Richtung?
Wollen wir das so bauen?
Wollen wir das so bauen?
Also das, was uns introvertierter Entwickler eigentlich die ganze Zeit so genervt hat, dass man sich immer wieder mit so doofen Menschen unterhalten muss und da eigentlich überhaupt gar keinen Bock drauf haben, das wird jetzt die wichtigste Fähigkeit, die wir haben.
I can do that.
I have also never had to go with someone to get into the software.
I was good and now I have a few of them.
They do what I say.
It's all super.
But that just this with other people interhalten, gemeinsam conflict through life, Entscheidungen treffen, and others widerlegen or bestätigen, and the opinion of it is now exactly the only thing that is left.
I think that the Stereotype could not ironically as by Entwicklern.
Why would you become a Entwickler?
Because it's a totally stereotype, what I say.
Of course, it's not everyone so.
But why would I become a Entwickler?
Because I think it's totally a charmant, to sit in a room where I'm sitting, where I'm going to sit and make it, what I want.
My standard example is, when I was 10 or so, I bought a VC20.
What was the first thing I was going to tip, when he was on the TV, was I was going to tip.
Marcus Andrezak, of course.
And then the name of Marcus Andretz, that was a TV.
My mother comes in and says, hey, you're in TV.
I said, wow, that's just magic.
Stimmt, I'm in TV.
Boom, boom, I'm a developer geworden.
Scheiß drauf.
And I think really, I have a control area where things work, and I'm the Herrscher.
And that works even better.
je weniger Menschen involviert sind.
In dem Moment, wo Menschen da sind, deswegen sind Stand-Ups morgen doof und so weiter, ist das alles Horror, weil jetzt bricht diese Matrix, die ich mir perfekt gebaut habe, zusammen und wird eine Singularität, wo plötzlich Menschen eine Rolle spielen.
Horror.
Was bleibt jetzt?
Das sind diese 10%, die ich vorher gehasst habe.
Die bleiben jetzt übrig.
Wo ich mich vorher ...
Keine Ahnung, von den 40 Stunden im besten Falle, die ich bei der Arbeit verbringe, immerhin vielleicht zu 15, 20 Stunden Coding kamen, die sind jetzt weg, die übernimmt die.
Die restliche Zeit wird mit drüber reden, out of all things verbracht, what the heck, das ist doch nicht, warum ich diesen Beruf gewählt habe.
Ich glaube, das ist wirklich so, mit anderen Worten im großen Bild, Entwickler entwickeln AI.
Der erste Beruf, den es trifft, ist Entwickler.
And what happens is that it takes the development away and what remains is the talk.
Yes, exactly.
There's a higher up.
Yes, exactly.
And then all these beautiful things, this not invented here.
Yes, I see what you've done here, but I find it not so nice.
I make it just myself.
All that was...
the most important part of the development of life.
It's just all, it's just all, it's just all, it's just all gone.
Sorry.
Yeah.
So, and how, also, what, you have said, and that's also my experience, also, I have two, I don't know how that's for you.
My observation is that very seldom AI and Unternehmen from Tech get driven, lustily.
Also, in the most companies, with which I'm in the way, Marketing, Sales, wem auch immer getrieben, keine Ahnung.
Und dann stehe ich immer überrascht da und sage, huch, die machen das doch und die wollen das doch, aber so den Rest des Unternehmens mitnehmen nicht.
Vielleicht hat es mit den 10% zu tun und mit der Menschenkommunikation weiß ich alles nicht.
Also das würde mich interessieren, wie es bei euch ist.
Aber du hast was ganz Wichtiges gesagt, nämlich die zweite Beobachtung, die ich mache, ist, dass Leute oder Unternehmen nur Fortschritte machen, wenn sie Lean-in machen und sagen, ...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
Also das erste Mal, ja, Token müssen zur Verfügung stehen, das kriegt man noch gebacken, wenn man will, aber das andere ist ja mental und von der Arbeitsumgebung ein Riesenproblem, das man adressieren muss.
Ich glaube, ich fange mal ein bisschen weit außen an, weil das vielleicht noch wichtig ist.
Testo baut ja Messinstrumente und wir erzeugen schon immer unendlich viele Daten, also schon zu frühen elektronischen Zeiten.
Das heißt vielleicht...
was so ein bisschen der Vorteil, dass uns eben dieses Thema Daten und so nicht so ganz fremd war, auch wenn es nicht unbedingt direkt ein IT-Thema war, sondern einfach, weil es unsere Value Proposition ist als Firma.
Das heißt, es war von Anfang an schon irgendwie klar, wir müssen da was tun, sowohl aus der Führung als auch auf der operativen Ebene.
Es war ganz klar, dass...
Das können wir nicht ignorieren.
Auch schon zu Zeiten, wo die Ergebnisse wirklich noch fragwürdig waren.
Also wo du noch Halluzinationen hattest in großem Volumen, wo du wirklich nicht sicher sein konntest, was dir die AI da zurückgibt.
Wo du teilweise selber nochmal reviewen musstest.
Also ich sag mal jetzt JGBT 3.5 oder noch früher.
Aber auch da schon gab es die Prämisse, Leute, diesen Zug können wir nicht verpassen.
Das ist so unglaublich wichtig.
That has also led to that we can be in a certain way experimentally experimentally be able to be able to do it.
And what we have done is to look at us regularly about it.
That's super important.
We make a lot more Hackathons than before.
We make a lot more Regal-Termine than before.
Where we also sometimes socializing, where we pack a lot of moments together.
um uns darüber auszutauschen.
Die driften dann auch schon mal ins Philosophische ab oder so.
Aber der Austausch an sich ist so wichtig darüber.
Wir haben halt Leute, die sind extrem weit vorne und wir haben Leute, die tun sich gerade noch schwer damit.
Wie baue ich mir denn jetzt erstmal meine erste Cloud-MD?
Das ist ganz, ganz normal.
Und wir versuchen möglichst viele Momente zu schaffen.
Und ja, das bedeutet vielleicht auch so ein Stück weit, dass...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
We have really found that we are at some point mass and longer geworden.
Also our metrics are okay.
Our quality metrics.
Bisher is this effect not yet.
That would be said.
It doesn't mean that it will come.
I have a little bit of concern about the employees.
That is not like a high-end or so, but I am aware of it.
I work to do this.
Because it's so exciting and so fascinating.
And I have the concern that the developers do.
It's amazing, what's happening.
And that makes me a little bit of a fear.
I have a lot of money, but I will not know how many people are at the end of the week.
Because that's so crazy.
That's almost a official topic.
Both Simon Willison and Steve Yeager have their own words.
One says, I can do everything and I can even go to a dog, when I do that, but at 11 am I am trained.
And the other says, we are AI vampires, we are all over these 100 things.
Boris Cheney says it a little bit differently.
He says that it's officially an ADHS-Centwicklungsstil, which he has with 10 terminals on to every time and, I don't know, 30 commits am day or whatever he does.
Or 30 PRs even.
And you say that it's a responsibility.
That's one thing.
But I think that the responsibility has also an economic factor.
We have to look at all.
I'm going to see that in private.
Je mehr mich das begeistert, umso mehr muss ich strikt für den Rhythmus achten oder sorgen, dass ich Festtermine habe, wo ich laufen gehe, spazieren gehe, damit ich nicht da reingesogen werde, weil es halt im Prinzip dieselben Suchtverhalten ermöglicht wie Social Media, weil du immer den direkten Sense of Achievement hast.
Jeder prompt liefert ein Ergebnis für dich noch weiter rein.
And that, what's on the Chat-Windows, happens at the end, also in the Entwicklungsumgebung.
That's not explizit.
But the next Befehl, which you have done, the result is just five seconds away.
And I think that we have to, as the society has better done social media, probably as a company already done, that we can create some rhythms that maybe super traditional are.
Also, I have in Sweden worked.
That was me totally impressed.
I was like, I was like, I was like, I was like, I was like, what's happening here?
Why is the clock?
So, lunch together.
That's cool.
At 11.30pm came the next clock.
What's happening now?
Yeah, now we go to the dinner together, these all lunch restaurants that it's here.
Oh yeah, so, you have to be quiet, you come back, 15.30pm next clock.
What's going on?
Coffee drink together.
So on the day came another extra clock, because today is coffee drink um 16.00pm again or something.
It was then verschoben, because now with Canelbulle and so.
And I, also, that was for me absurd in the often times.
Solche Sachen wären wahrscheinlich heute relativ sinnvoll, um die Leute aus diesem Always-on rauszukriegen.
Absolut.
Dieses Closure-Dopamin, wie du beschrieben hast, immer wieder nochmal und nochmal was geschafft und nochmal was geschafft und ich komme dem immer näher.
Das ist wie TikTok für Coder im Prinzip.
Du bist nur noch am Doom-Coden, weil es einfach auch so, ich schreibe auf einmal Swift-Code, ich schreibe auf einmal Rust.
And they work.
And I write tests for it.
And I have my workbook so developed.
And it makes so fantastic fun.
But the Rechnung comes.
And that's also very bad.
I'm not a big fan of the theoretical.
Aber du bist natürlich ein Role Model.
Wenn du deine Mittagspause codest oder wenn du bis 18 Uhr noch irgendwo rum codest und die Leute sehen irgendwelche Commits von dir bis um 21 Uhr, dann hat das einen Effekt.
Also erstmal legitimierst du das und dann gibt es vielleicht sogar noch welche, die fühlen sich davon angesprochen und sagen, ich darf ja nicht früher gehen als der Chef oder so.
Or am Wochenende generate Dokumente in Slack.
Yeah, genau.
Mail schreiben.
Weil es geht und so.
Das ist ja nicht schlimm.
Ich habe ja gar nicht gearbeitet.
Nee, das ist ganz schön krass.
Und das finde ich gut, wie du sagst.
Man muss sich da im Prinzip viel härter als früher da einen Tagesablauf bauen mit einer relativ eisernen Disziplin.
Da kann man sich ein Programm verschreiben lassen.
But that's totally important.
The program is still a way to go.
Yeah, yeah, exactly.
Notification, that's we learned, that Notification can go on.
Yeah.
Next question.
In my theory is that as a whole, the whole process of the process of a small road, a bottleneck and we took away.
Am Ende ist es doch nur Code generieren.
Früher war mein Standardbeispiel immer so von schlauer Mensch steht unter der Dusche, hat geile Idee.
Bis das wird wirklich was und kommt beim Kunden an, ist eine elende Strecke.
Und egal was wir bauen, meine Rechnung war immer, weil das ist so ein Produktmanager-Problem, wenn das Coden nur zwei Wochen dauert oder keine Ahnung.
20-Mann-Tage.
Wieso brauchen wir dann vier Monate, bis das Ding auf dem Markt ist?
Naja, weil wir eine Unmenge Wertschöpfung drumherum haben, bis das Ding zum Kunden kann und so.
Und selbst wenn wir nur Softwareentwicklung angucken, ist halt Code-Generierung auch nur ein kleiner Teil davon.
Das heißt, das kriegen wir geschenkt.
Und jetzt sagt uns Theory of Constraints, dass wenn wir das geschenkt kriegen, poppen.
downstream und upstream andere Bottlenecks ab.
Wahrscheinlich ist das sowas, worüber alle gerade reden.
Auf der einen Seite Taste und sozusagen was bleibt denn bestehen von dem ganzen Kram, den wir hier bauen?
Was schmeiße ich weg?
Da wäre übrigens eine Metapher, die mir noch eingefallen ist, Analogfotografie zu Digitalfotografie.
digital and with the handys you make you just 1000 photos, you look at it again, you can see it again, you can see it again, you can see it again, we can see it again, we can see it again, what we hold or not, because it's not bad.
That's a question, how do you get it with something else?
The other thing is, underneath comes the famous PRs, which could be everywhere.
I don't know if that's for you so, but what are you things that are now on the way to make a lot of code-generated?
And just that?
This Bottleneck-Code production and building systems, we're at the level of the way that this is not a bottleneck anymore, but now everything is a different bottleneck.
I have the problem that I have understood?
Now I have to talk about the customer.
How do I get a date?
This whole chain, are my annumations?
How do I get the information?
All this, then I have a decision to really invest in.
I say, okay, I invest in, I'm going to go now.
How fast I can meet the decisions and how often I can meet the number of weeks when the code is done in the day.
Also, like you said, when you say that when you see that as a whole Lieferkette, then is of course a part of the Kette now rasend schnell.
That was before the big bottleneck was.
But it doesn't mean that the Lieferkette will go through the way.
But in the same way, now pop the other things up and I feel like...
Super Analogie mit den Fotografien.
Ich erlebe gerade im Moment so einen Reflex.
Ja, dann bauen wir halt alles.
Bauen ist ja billig.
Dann holen wir halt die Schrotflinte raus und gucken, welches Schrot trifft.
Also das ist halt dann ein riesen Kegel oder eine riesengroße Oberfläche, die wir jetzt einfach beschießen mit Lösungen und wir gucken einfach, was stecken bleibt.
Wo man da glaube ich Aus meiner Sicht macht man da den Bock zum Gärtner, weil das ganze Thema Betriebskosten, das ganze Thema Maintenance, das ganze Thema Governance, Compliance, das vervielfältigt sich ja exponentiell genauso auch mit diesem Thema.
Und das kannst du nicht einfach an Agenten geben oder so, sondern das musst du trotzdem noch machen.
Und das, also da sehe ich halt im Moment, sehe ich das als natürliches Verhalten, wenn wir so viel bauen können, dann lass uns doch einfach so viel bauen.
Und du You riskierst aber deine Identität, du riskierst deinen Brand, du riskierst, je nachdem wie du es machst, man kann das natürlich auch gut machen, aber du riskierst deinen Überblick und ich glaube, wir brauchen eher die Fähigkeit, alles davor auch so zu beschleunigen, dass wir wirklich immer noch sagen können, okay, und das gehört zu unserer Identität, das gehört zu unserem Brand, das wollen wir tun und das wollen wir alles nicht tun.
Aber die Fähigkeit, I know very few companies that have already been designed for that.
So for the companies that can't be done.
I have already seen the experience in the environment.
You have the dream.
You have a product manager that we have promised.
They run around and build the prototypes like the Wahnsinn.
And they meet each day two hours and show the prototypes.
Actually, they all have a prototype, because they are great now.
They would like to produce all of them.
So, and then you can even go all this code.
And I think that the main bottleneck that we don't have to be aware of is the very, very vorne, the customer.
And then I'm going to give a example.
As I worked at the under-under-under company from eBay, the biggest problem that we had then was Amazon is better than eBay.
And why is Amazon better than eBay?
Because you have to wait for the end of the auction, because you have to wait for the first to buy.
Also Amazon has to do the first to buy.
And that's technically no problem, a sofort shop to build.
That problem you have, is to tell the customer that you're just someone else.
You have explained the customer that a auction for the famous PETS-Spender is the best in the world.
And that all long tail is.
The customer makes the experience that things just like that, super nice.
So we can also.
And the customer says, what?
The guy who told me that I told you that I'm going to say that I'm going to say that I'm going to say that I can't buy it anymore.
I don't think I can buy it anymore.
We have been waiting for years to put it on the position to be able to put it on the way.
A lot of years has been waiting for the customer to buy it.
It's not a matter of how many product managers with how many prototypes are in the field.
That would have been solved.
Kaum war das gelöst, kam es nichts, nämlich dass wir sagen, hey, Kleinanzeigen können wir auch nicht.
What the hell?
Kleinanzeigen?
Das ist ja nochmals ganz anders.
Das seid auch ihr, habt ihr sie noch alle?
Wieder Jahre, fünf Jahre, bis das die Leute begriffen haben, dass Kleinanzeigen beim selben Unternehmen jetzt irgendwie auch funktionieren und so.
Und ich glaube, das bedenkt man nicht bei diesem, wir wollen das alles bauen.
Also die philosophische Frage, ohne dass ich AI und sowas kleinreden will, ist ja wirklich, We are not going to be a company for something very, very, very certain.
And if that's too big, we don't understand the customers anymore, who they are.
And we don't need a sub-profit like Gore-Tex.
How big does a sub-profit be?
I don't know.
I have the answer to that.
But in how far we deliver is really, really, really the problem?
That can, I believe, very, very different from a company.
Even if we always...
Just because it's bottlenecks, it's always the actual bottlenecks.
And at the moment was software, also code producing, was a huge bottleneck.
And yes, that changes everything.
But is that the bottleneck now?
No, probably not.
Deine Produktionsgeschwindigkeit ist einfach limitiert durch die Adaptionsgeschwindigkeit deiner Kunden.
Du kannst so und so viel liefern, du kannst riesen Produkte bauen, solange dein Kunde oder der Benutzer, nicht immer von Kunden reden, aber der Benutzer nicht in der Lage ist, das zu adaptieren, weil er ja selber auch limitiert ist, bringt dir das also überhaupt nichts.
Was dann jetzt passiert ist, was ich gerade so ein bisschen erlebe an mehreren Stellen, also auch in anderen Firmen, ist, dass man jetzt versucht, ...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
Und der funktioniert auch ganz anders und ich merke gar nicht, wie ich da reinrutsche und auf einmal vielleicht ein Versprechen abgegeben habe, Verfügbarkeitsversprechen, Leistungsversprechen, Wertversprechen, was ich vielleicht gar nicht langfristig halten will.
Also das sehe ich als großes Risiko an.
Und die Frage ist immer, warum sollte ich es können?
Also wenn du dir ganz, ganz großen Unternehmen anguckst, wie viele Märkte haben die?
Also ein riesen Minderwertigkeitskomplex von Google ist ja, dass sie sagen so, We have always in the Kern a model that has been carried out.
Apple, also Spotify, first of all.
Then they have a podcast or so.
But at the end they say, no, it's all the same same.
And we're always talking about that there's no way of advertising.
And we're not doing that.
But to start a new way of creating a new way of creating a new way of creating a new way of creating a new way of creating a new way of creating a new way of creating a new way of creating a new way of creating a new way of creating a new way of creating a new way of creating a new way of creating a new way of creating a new way of creating a new way of creating a new way of creating a new way of creating a new way of creating a new way of creating a new way of creating a new way of creating a new way of creating a new way of creating a new way of creating a new way of creating a new way of creating a new way of creating a new way of creating a new way of creating a new way That's what happens in the work.
I can now research.
I could write a program, which is my free income.
All possible could I do it, but if I do it or not, I have to ask myself to ask myself.
I come to the idea of the thought, that probably the whole field changes, but probably not so, as we think.
We're going to move somewhere else.
Internet has changed the way to the product manager, the product manager, the it's been today, it was not.
We know it probably not.
And at the moment, the theory, I don't think, is that all the field are going to be further.
All the people are downstream and upstream.
The product manager is strategical and taktical and coded a little.
And I think that what we are currently seeing as technical possibility is that it can be done, but it will be given by what the person can and where he feels comfortable and where he feels competent.
And my experience is that every time you always find Generalists, with which everything is gathered.
Also like a time of product management, you can suddenly be a generalist, from data analysis to strategic thinking, to design.
or even make a decision.
Also this whole weird thing about skills.
And see there!
There are not many people who are generalists.
And I think that's what's really quickly crystallized, that this is a cool wish, which is the one that can understand economically.
But that's not the way the human nature is not so, that 80% of the population are generalists.
Aus bestem Grund.
Also wahrscheinlich ist das sogar physikalisch fast begründet, wie die frontale Kortex bei uns aufgebaut ist.
Ich weiß es nicht wirklich.
Aber wir haben noch nie gesehen, dass sich Generalisten langfristig durchsetzen.
Langfristig kam immer die Peitsche zurück von der Spezialisierung.
Ja, im Grunde genommen zwingt uns, also gerade Entwickler, aber ich denke, das wird sich in anderen Berufsfeldern auch so zeigen, im Grunde genommen zwingt uns die KI oder AI.
two things on, which we actually don't like to do.
One is to talk to each other, to deal with conflict, to create conflict, to create Entscheidions.
And the other is to say to the whole option, which suddenly arise, which only because they are there, still not a good idea.
And these two possibilities now to build out, that is, I think, the big challenge of all organizations.
Also there really good to be in.
The, the, the, the, the, the, the, the, the, the, the, the, the, the problem to find, for what you need a solution.
And that also with more people there working together.
That is, I believe, our biggest challenge.
You have to ask another second part of the question, how we do this now.
I have, I believe, yesterday on LinkedIn posted, I had this Simon Willison...
Lenny's Podcasts gehört, die Folge fand ich wirklich fantastisch.
Und da ist mir überhaupt auch nochmal klar geworden, was diese verschiedenen Arten von Coding und Umgang mit der AI im Sinne von Software-Engineering bedeuten.
Und ich habe das jetzt für mich so ein bisschen so definiert, das Vibe-Coding folgt eigentlich eher so einem agilen Ansatz, weil ich extrem annahmengetrieben mich einfach mal darauf einlasse, was zu bauen und mal zu gucken, ob das stickt, ob I give it to my customers, I can tell you, look here, I have the idea, we have this in a 30-minute session, I have this done with Cloud Code or whatever, and we look at it just like that and see if it is something that can be worth.
That's what I can do now under Wipe Coding, also a less-fair Ansatz, where you are not so much in testing, in the Qualitätssicherung and so on.
Then there is the genetic engineering.
I would say that There are we, where we think about it.
Okay, how can I do this Coding as a nucleus take?
And everything, what I do do, CI-Pipeline, Testing, QA, Architecture, also have I an Architektur-Agent?
Have I maybe many different Front-End, Back-End Architektur-Agenten, the on certain Regents basing, the me concepts make.
What do I have for Koala?
Have I penible Koala?
Have I UX Koala?
What do I have for QA-Agent?
And I try to build code all the links and rechts of the network over AI.
Because I can now, as I can now PRs, I can now also free or able to do.
Or as soon as I can now software, I can now, I can now.
Agenten auch in der Lage sein, die Software zu bewerten und mir ein Feedback dafür zu geben, ob sie dann produktiv laufbar ist.
Also das verstehe ich unter Agentic Engineering.
Da arbeiten wir im Moment sehr viel mit, versuchen, da ist es noch besonders knifflig, weil ...
Da haben wir noch nicht so richtig den Team-Austausch gefunden.
Natürlich haben wir zentrale Repositories.
Die Leute packen ihre Agenten da rein.
Ich kann Agenten wiederverwenden.
Aber dann habe ich dann doch teilweise, wir sind Domänen aufgeteilt in Teams, hat das eine Team wieder einen anderen Bedarf als das andere Team.
Also da haben wir noch nicht so den roten Faden gefunden.
Und ganz on top ist für mich dann das sogenannte Thema Dark Factory.
Das heißt, ich sehe eigentlich keinen Code mehr.
I don't look at it anymore because I built this system, which has been built in the way, which ensures that behind there is really something needed and also qualitative and high-quality.
With people have been able to achieve it, from my view, to a large part.
But that must we have to achieve with the agent or with the AI.
And I think that it's worth it for everyone to listen to.
in the podcast and just the minute where the dark factory pattern was explained, there are two entscheidenden regels.
The first one is no Schreibcode, the second one is no Least in Code.
And what he describes is relatively exciting.
One of the companies that have tried it out of all companies on this planet is a company, which has a security software produced.
What they tried to do is, oh, Wenn ich Code sozusagen kostenlos jetzt generiert bekomme und ich aber vorher die Qualitätssicherung ins Codeschreiben reingebaut habe, dann könnte ich ja jetzt wieder rausziehen.
Wie sieht denn ein QA-Department aus, das keine Menschen hat?
Und was dann gebaut haben, ist Agenten, die User simulieren, die einfach die Software ausprobieren.
Die fragen die ganze Zeit um Access zu irgendwelchen Systemen und kriegen das geliefert.
Und dann guckt man, was geht schief und so weiter und so fort.
Und dann haben sie festgestellt, wir können das so hochkarätig gar nicht testen, weil die APIs alle great limited sind.
Jetzt müssen wir Slack und die ganzen Systeme, zu denen sie Zugang gewähren, simulieren.
Jetzt simulierst du nicht nur die Personen und alles, sondern sogar die Software, auf der das arbeitet und so weiter und so fort.
Ich will gar nicht sagen, dass das unbedingt erstrebenswert ist.
Wie gesagt, ich finde das zum Teil langweilig, wenn der Mensch raus ist.
Es hilft aber, das extrem anzugucken, um on good questions to come, I believe, and then on good solutions to come, which one another automation gives.
So, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so They make this software and they write a lot about it.
They have really already started with the concept of, hey, we don't have a code anymore and we don't write anymore.
And we need to build a system, what we have done before with people have achieved.
We have to build, we have to build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build, we build That's what they just did with the AI.
And they write a lot about it.
It's super interesting, how many models they had to go, to go through the borders, which they didn't know about it, which they didn't know about it.
It's a interesting story.
I can just tell you.
StrongDM.
There's also factory.strongdm.ai.
I will not spoil it, it's a weather report, but it's very exciting to read it.
What I also find as a last exciting thing, the answer is not just software to build, but qualitative higher software than before.
That's the answer.
The answer is better than before.
That's the answer.
It's a good thing.
We are now so far, we are always going to build from scratch.
We do not make any iterations in Code anymore.
Also, an iteration would say, I'm going to leave three and give five to five.
But that doesn't find anything more.
You're going to leave everything and throw everything out of a sample of prompts with your agent.
Also, that's a super important thing.
A great experience for us.
And the second aspect, which I will call it, is...
Die Software verändert sich dauernd.
Du kannst deine Testsuit, wenn zu viel Zeit vergeht und wenn du den Prompt zu sehr veränderst, dann kann es sein, dass die Software, die beim dritten Mal gebaut wird, völlig anders aussieht als beim ersten Mal.
Trotzdem funktional ist und trotzdem performant.
Das heißt, du kannst nicht davon ausgehen, dass du eine Testsuit einmal erzeugst, die irgendwo ablegst und dann regelmäßig gegen die Software rennen lässt.
Vor allem nicht auf der Unit-Ebene.
You have to go with that, that the software has four or five times a completely different look like before.
And that's also a challenge, which we haven't been given to.
Also, we have to go with that, how we can do it.
What are the reactions to the moment in the development community?
In the development community?
At the moment it's all...
It's at least my point.
My colleagues and colleagues should be corrected.
But at the moment it's all...
crazy, interesting.
Das geht auch und das geht auch und das kann man auch.
Also ich habe natürlich den Bias, dass ich selber auch dem Ganzen mittlerweile nicht mehr so skeptisch gegenüber eingestellt bin.
Ich habe den Bias, dass ich super gerne mit den Leuten auch rumtüftle.
Es gibt mit Sicherheit auch Leute, die, wie ich eingangs erwähnt habe, die halt sagen, nee, ich habe da keinen Bock drauf.
Sorry, aber das sind Tech-Buros, das ist Energie.
Absolut for the Earth and so on.
And I have to give it to the rest.
But at the moment I have the impression that this Neugierde, which is still not yet again.
It is the end of the day.
I'll leave it again.
But I have another question.
You write a little about the system that you built.
I think you put it even on GitHub.
Yeah, Codeberg.
a breadth of knowledge model that I can learn.
That's super important.
And that's my next project.
That's my next project, on the way I work.
You can find everything about me on OliverSchwarz.info.
There is my best one.
There is my LinkedIn Account.
There is my Codework Account.
I write a little bit about it.
But there you find me on the way.
And then of LinkedIn.
I have been very much for...
I'm not at X, not at Facebook, Meta, Instagram, everything away from the reasons.
I'm on Mastodon.
And also the Mastodon-Account you can find on oliverschwarz.info.
I'm always happy to ask questions.
And I live with these two hearts in the chest, where I say, actually is that pretty and doof.
But this technical fascination has also been caught.
And I can't get the whole hate or the harsh criticism of the criticism.
I can't win.
It's a little bit.
The button is pressed.
No one takes no more.
I would have to do a rational conclusion that can be a little bit.
I think with all the doubt, the one has...
It's just to meet the person who understands and understands.
If I was a big fan or a big fan, I would just understand it, so I could do it in a cool Tai Chi move to overcome it.
But without a understanding of it is absurd.
Just that stuff, what happens, that not to understand and just being there, will just work out of it.
Um es einzugrenzen, wird man es verstehen müssen.
Da hilft nichts.
Ein paar können auf die Rente warten und sagen, geht an mir vorbei und so trifft dann nur meine Kinder.
Aber gerade unsere Kinder, wird es alle betreffen, was das jetzt gerade alles macht.
Man kann es denen überlassen.
Ich glaube aber nicht, dass das die aufgeklärtere Haltung ist.
Es gibt da einen wunderschönen, also wirklich den schönsten Artikel, den ich jemals I have to read the subject of Tech from Katrin Parsich.
I don't know the title anymore.
I think it's called Phasen of Technological Critics or so.
Where she goes back in the history and goes back and goes back and a little prosaic, but is totally interesting to read.
As the first gas-laternes in Paris were built or so, the Bauern, the aborment they said, we will be overwashed or we will be overwashed.
Und diese Reaktionen zu, sage ich mal, technologischem Fortschritt, die werden da in einem Artikel schön zusammengefasst.
Und das ist so wunderschön und das sieht man halt jetzt auch wieder.
Sorry.
Also ich habe noch ein gar nicht so weit zurückliegendes Beispiel.
Das habe ich persönlich noch mitgekriegt und da konnte ich auch schon wieder nicht mehr entwickeln und so.
Aber ich habe, als ich bei Mobile.de gearbeitet, getrieben wie der Teufel, dass wir diesen ganzen CICD-Kram machen.
Und das war ja am Anfang, war das ja...
It's the same as today.
Everything is the same.
That doesn't work.
And if you're in Greenfield and just in the game, but we have a real software, we're going to have everything kaputt, if we just think about it.
And then we've been doing it a half year or so.
And then you're going to tell it's like, like today, come under Bottlenecks.
Oh, Qualitätssicherung.
And now, without automated testing and testbed, it doesn't matter.
The manual QA can't forget.
What?
What do we do with the people?
But it was just a devil's thing, as you just described it.
And it's all been released as a possibility.
What you read, it was first a few Blog posts and you have all been like a spinnaker and the idiots.
That's exactly what we're talking about Simon Winsons, Boris Cherny, Steve Yeggis, where they say, ha ha ha, for them, that's not in real life.
And I swear, that's what we're going to do all of them, regardless of we like it or not.
But that's exactly the movement.
And everyone who does not know will say, you are the guy who said, 10 releases am day is completely impossible in the serious way.
It's really not long ago, that we were in the same situation and said, it doesn't go.
Yes.
Yes.
Yes, exactly.
As we did that, I know how today, we had a Christmas party and the people said, what was the best we did this year?
CICD Pipeline, super cool.
Alle, die das schlimm fand und gekämpft haben, hinterher war es das Beste, was wir im ganzen Jahr gemacht haben.
Und das ist ja eigentlich schön.
Warum ist das schön?
Weil Bottlenecks, die weggehen, sind immer cool.
Das merkt jeder, der mal ein Bottleneck war.
Es ist nicht schön, ein Bottleneck zu sein.
Ja, ist richtig.
Ja, und ich kann dir da nur zustimmen.
Das ist genau mein Eindruck.
Und das wird hier auch Yeah, we're going to, also it's going to be a thing that doesn't work.
And it's going to be a thing that will happen.
Everything, everything.
Absolutely.
But we're going to be a thing that will not work.
And let's do it then.
Then we're going to be a thing.
Oh, yeah, very nice.
What can you...
Also wenn man dein System findet, ich gebe das in den Show Notes an, was kann man damit machen, was hilft dir Feedback oder ist es eher lästig, wie ist das irgendwie?
Also gerne jederzeit einsetzen, wenn, dann auch gerne Feedback.
Ich habe mittlerweile eine Installation MD dazu gepackt, die ist nicht für den Menschen, sondern für Claude.
Ich versuche es gerade auf einem Gemma.
4B ans Laufen zu bringen.
Also ich versuche gerade irgendwie mit lokalen Modellen daran zu kommen.
Bin da aber noch nicht so weit, wie ich da eigentlich gerne sein würde.
Aber das Allerbeste ist die Installation MD an Klo zu füttern.
Dann würde er das System im Prinzip von Null aus hochfahren und dann einfach mal damit rumprobieren.
Man kann es ja im Ordner machen, kann man wieder löschen.
Also es passiert nichts Schlimmes.
Ja.
Ich wäre schon interessiert über Feedback.
On the other hand, I'm not a service.
It's open source, it's MIT-Litends, so please go ahead and try it.
But I'm not a service.
There's no commercial interest in it.
And if someone finds it, the cool thing.
Exactly.
Mega, Oli, thank you.
You took a lot more time, as I asked if you had it.
That was super exciting.
Yeah, but...
...and for the real life.
And...
...toll.
Also, I look at the system...
...and come back and see also solutions...
...for me.
That's super cool.
But thank you for the time and the time you have given us...
...and all the good and until soon.
Dankeschön.
Ja, danke dir für die Einladung.
Danke.
Es war echt spannend.
Zwei Stunden jetzt.
Super, super, super.
Hat mir auch voll Spaß gemacht.
Und wir sehen uns auf jeden Fall in den nächsten Tagen irgendwie.
Cool.
Bis bald.
Dankeschön.
Tschüss.
Tschüss.
of Marcus Andrepack.