KI schläft nicht !

Transcript

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

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

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

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

00:00:24Designed to provoke thought, make you smile, and above all, to engage in conversation.

00:00:29A warm welcome to Thinkdifferent, Think AI.

00:00:37And today we're once again without a guest, but I promise you, that will change soon.

00:00:43Jens is here with me. Hi Jens, nice to have you here.

00:00:46Hi Mark, nice to be here and nice to have you here.

00:00:50Yes, that's great, right? We'll do this for ten minutes and then hang up.

00:00:54No, don’t worry.

00:00:55Jens, I have to say, I was almost blown away recently.

00:00:59I opened LinkedIn, and as some may know, I spend some time there.

00:01:03And then I saw an article by you.

00:01:05And it stuck with me because it had a visualization that just captivated me.

00:01:13I mean, I don’t know how it is for you. You go on social media and you either find bad photos or good photos.

00:01:19You find bad thumbnails or good thumbnails.

00:01:21And I think we're going to talk a bit about AI images at the beginning.

00:01:26Would you briefly tell us why the article was made and especially,

00:01:29where that cool image came from and what it showed and how that relates to our episode?

00:01:33Yes, the article, the original was taken down, so I need to think about it again.

00:01:39That was, I believe, something about agents, it was about agent relationships.

00:01:43It was about the fact that, in principle, the topic is becoming increasingly important that we

00:01:48do not have to see agents as deterministic little programs that just process something,

00:01:58but that, in principle, in this whole context, as companies, as people,

00:02:04situational situations arise, where agents play a role

00:02:11and it must be seen more like a relationship.

00:02:14It must be seen like that. I'm not talking about emotions, all that other nonsense, but I'm saying,

00:02:19yes, here there is an emotion. That’s nonsense, I find it beautiful. But it’s really about

00:02:24actually saying that we need to establish this interaction, these workflows that we set up,

00:02:30we need to stand like a relationship. And not like a, I start a cron job somewhere,

00:02:35that then runs, but this job, what has to run, needs

00:02:40results, goals, to which it can work, it must also have the ability

00:02:47to decide things based on certain authorities that are given or taken by someone else.

00:02:53That also needs to be known. So, that was the article printer rotor,

00:02:58I don't want to go on for too long. And I had such an illustrative style in mind. Normally, I like

00:03:02to do it in a pixel retro art style, because I also like to reminisce about my old C64 gaming past

00:03:09when I post something. But this time, I've opted for something different.

00:03:12That’s where our podcast cover also comes from, it must be said.

00:03:14Our podcast cover was significantly shaped by you, and I have to think every time I see

00:03:19that in our podcast episode, mentioned Manage-Mentions.

00:03:24But correct.

00:03:25Correct.

00:03:26That’s also kind of the point of the Fekter, but in principle it works

00:03:29now, I've also noticed it with one or another post, it doesn’t work

00:03:33that well, the style, when I want to make illustrations or the content

00:03:38of the article is not only supported by a graphic but should actually tell a story,

00:03:43but an illustration, a real

00:03:47infographic, then this pixel style doesn’t work well, so I’ve

00:03:51oriented myself a bit differently, looked at what else is out there. And

00:03:54indeed, I like to do a combination with Gemini,

00:04:00so that there are also bananas in the background, because that was basically the first AI that made

00:04:05waves in the last weeks and months regarding infographics and simply produced good results.

00:04:11And text consistency, right?

00:04:13And there is text consistency, right? Had to think about music for a second.

00:04:18ChatGPT has that too, actually which graphic model I used in the chat...

00:04:22But that's later.

00:04:23But later, so...

00:04:24Do you know which model?

00:04:25Yes, who is that?

00:04:26Oh God, ChatGPT.

00:04:28I can look it up while you talk.

00:04:30Yeah, we can do live research while you talk for a few seconds,

00:04:33I can quickly check it out.

00:04:34And so that was, I took that, so in principle worked with Nanobanana and then mostly went through Manus.

00:04:43Which also takes like Nanobanana but understands my context a bit better, what I actually want to do, and additionally had a functionality that GeminiZone hasn’t yet brought.

00:04:56So if Manus uses Nano-Banana to regenerate an image,

00:05:03it should then be placed on the own canvas, which is basically an artifact in the Manus chat.

00:05:08And I can then edit text fields on this graphic individually,

00:05:13I can go in with Edit Text and edit these text fields.

00:05:17This is of course great, because there are still not too many

00:05:21typos that the AI produces, but it is mainly about saying,

00:05:25Now there is text on it, and I often do it that I just take the article text,

00:05:30just paste it into the AI and tell the AI to make an infographic out of it.

00:05:34Then it can still be the case that some

00:05:37text blocks, some labels on the right or left, top or bottom, are not so on-point and I would like to rewrite them.

00:05:43And if I had to prompt all of that, then you have again all this effort,

00:05:47I would have to say roughly which main points exactly,

00:05:50everything gets rendered again, and it happens that it looks different again

00:05:53And through this text editing functionality, that was always the level for me,

00:05:57where I could intervene and change it.

00:06:00That has changed a bit due to what happened this week, right before the weekend.

00:06:06Quite good, before we name that, so that I can simply fulfill my duty to inform.

00:06:12The model has the really great name GPT Image 1.5.

00:06:19I find the Nano-Banana prettier. Because those are always the Minions, Banana, but that’s a different topic.

00:06:27Say, what was the name of the OpenAI graphic model in the beginning, that everyone celebrated?

00:06:32I can't remember.

00:06:33So it wasn't Dolly. What was it again?

00:06:35No, Dolly, but wait, no, what? No, it's similar.

00:06:37Dolly.

00:06:38Dolly, Dolly.

00:06:39Oh yes, Dolly. Now I have, I have Dolly.

00:06:41Dolly.

00:06:42Yes, that was my bad pronunciation.

00:06:43Yes, okay, but that was the model from OpenAI back then.

00:06:47It seems to have disappeared into oblivion, right?

00:06:51Perhaps with good reason, but that’s another topic.

00:06:56One has already said that we had

00:06:58recently an episode called Mythos Entropic, Entropic Mythos.

00:07:02And now suddenly, neither Mythos came out,

00:07:05nor a whole lot of other stuff, new models,

00:07:08Claude, what was it, 4.7, Opus 4.7, something like that.

00:07:12It’s quite funny.

00:07:14But we got into it because I saw a link in the post

00:07:17because you were using another system from the house of Anthropic, which then came out on a Friday

00:07:23and has certainly already caused problems for other companies.

00:07:28What is it about?

00:07:29Would you like to tell us more about it?

00:07:30Yes, sure.

00:07:31So I'm using it now for the new articles.

00:07:34This is the new Cloud Design, one has to say about Anthropic at the moment, they are bringing

00:07:41more features, models, and other software solutions are currently being released more than other people change their

00:07:48underwear.

00:07:49It's really crazy what's happening right now, isn’t it?

00:07:51It's like the beginning of a new month.

00:07:54Exactly.

00:07:55Some do it once a month, but with, well, actually with Tralfik you get the feeling that they

00:07:59release something every hour because they are generating a lot themselves with

00:08:02AI.

00:08:03They've mentioned this a few times already.

00:08:04What they released last Friday, I mean, here in the middle of April,

00:08:10is the topic of cloud design. That means, alongside cloud work, cloud code, cloud chat, however

00:08:17they all call it, there is now also cloud design. I don't have it in my application

00:08:22that I have installed as a headline on my computer,

00:08:26but via the browser, you can access it again if they

00:08:28register you through the browser. I don't have it through the app, but in the browser.

00:08:33But you need to have the pro subscription, the one for 20 dollars or something like that.

00:08:39Maybe, I'm not under Max, so from that side.

00:08:43It's there; it's like people always ask me about the iPhone,

00:08:47when did that start working, does this still work on my iPhone?

00:08:51If you buy the latest iPhone every time, then every time it's just a

00:08:55surprising moment of new features, always manageable,

00:08:58but you lose track of time and space,

00:09:00You actually experience, when I realize on which devices it actually works and sometimes you pick up an iPhone and think, how is this not working? What’s going on? Is it broken?

00:09:09No, no, two years old. What a pity.

00:09:11Okay.

00:09:12Now, not all of our listeners are like some Roman emperor lying in his chamber with grapes in the form of new models being brought to him for tasting.

00:09:25You don't have to describe how it looks here during our podcast recordings.

00:09:29Nonetheless. So, as I said, if you have the subscription,

00:09:34it's a kind of, I believe this is also the pro subscription that you can choose,

00:09:37below that there’s not just the free subscription, then you have this

00:09:40feature now, cloud design. Just as you hinted earlier,

00:09:44it made a bit of waves right at its release, it’s

00:09:47already caused some ripples in the social networks of this world that

00:09:51there is a new player that will definitely shake up the design world again.

00:09:55Because there have been many new tools during this time. Google Stitch has been

00:10:02mentioned, I think it appeared two or three weeks ago, that it's really a

00:10:08nice tool if you want to create things, whether they are wireframes, posters,

00:10:12or any other things, there are currently many topics coming up and

00:10:16cloud design is another one. If you just look at the examples of what you can

00:10:22do, from a purely wireframe design to prototypes with functional,

00:10:29animations, everything is possible, configuration options. So I can also say alongside the design I

00:10:36want to see, let me directly change such a second view

00:10:40where I can actually have sliders to change corner radii, change step sizes

00:10:46if I want to, to change the speed of the animation that you

00:10:51apply some lens flare effects or joke effects to the images on the website, so

00:10:56highly interactive, incredibly good, I think, now to build prototypes directly, when you

00:11:01do any things, have the first ideas, can build prototypes and it's extremely good also in the

00:11:06application, when you have existing design files, have design systems, then you can also

00:11:13Yes, you can also use, like from Figma and so on.

00:11:16Everything like that, you can, Figma, if you are directly attacking Figma, actually, because

00:11:21they directly offer to import the Figma files to learn from them, then you can

00:11:26also simply use a Chrome plugin then to pull web content if you particularly

00:11:31like it, simply copy it into the prompting, to then,

00:11:37I don't know, you'll get a good graphic somewhere or a navigation

00:11:40quite well, then you screenshot that, take it over, you can simply paste it in, have

00:11:44basically the theme directly within your design system or cloud in this case, take it

00:11:49just open it up and take it over.

00:11:51I mean, I'm not the graphic designer under the gentleman.

00:11:54I don't actually want to say whether the term graphic designer is correctly chosen in this context,

00:11:59if someone is aware of their profession.

00:12:03if someone feels attacked, it is unintentional.

00:12:06But one thing I'm sensing is, I definitely maybe have a

00:12:10feel for what I like, but I have no feel for how to create an interface

00:12:14in a way that meets my own standards. So I definitely need

00:12:18people with a knack for that and it was then to assess, is this

00:12:23good, is this bad. I mean, we've been making apps for many years, so at some point

00:12:28you might develop a feel, still I can't do it myself. And now

00:12:31I've only ever dealt with Figma on the sidelines, and when I saw that

00:12:35I left on Friday, somehow, no idea, didn't have time on Friday, then on the

00:12:41weekend I read about the other one, took a look at it, and then thought, damn it, the dynamics

00:12:48that it offers you, the openness it offers you, the possibility, you'll already

00:12:53just said, I reference sources and I tell him something and then I get suggestions

00:12:59and can engage with him, it's certainly not just something we easily

00:13:03We've mentioned a few times that you don't have to ride on the rome, there are definitely others.

00:13:07The challenge makes, but I can also do something like posters with it.

00:13:11The classical world of design might come to my mind now.

00:13:14I can create layouts with it. In the classical world, PowerPoint comes to mind.

00:13:18And now I have the opportunity to create something like, I think they call it,

00:13:22do they call it design system with them?

00:13:24That's just off the top of my head, you put it here and then I have my values

00:13:27and you know, whatever, and whether I make posters or slides,

00:13:31suddenly cool things come out.

00:13:34I discussed this this morning with someone who thinks it’s like

00:13:36Kenver, it's like, no, I've never really understood Kenver, for me Kenver is

00:13:40also very, very powerful, what it offers, but with this, with these design things

00:13:46I currently feel like I'm getting to something faster and easier that I

00:13:52can also use later on.

00:13:53We mentioned it a bit during the intro, in the sense that

00:13:58just throw Knotkot at it and then he has something to work with

00:14:01and you basically get real functionalities and apply that in real apps and websites.

00:14:05and if you can’t see the explosion coming, then I thought to myself, okay, this is on one hand

00:14:09a change in the technological environment, but also a significant

00:14:13proof of what we have also discussed in previous episodes, I recall the episode

00:14:17with René, how humans must also remain open to changes in work tools.

00:14:27I mean, that doesn't mean everyone automatically has to use Cloth design stuff now.

00:14:32I mean, it’s just a matter of time before the next ones come out of the hole

00:14:37and maybe show something cool.

00:14:39But it shows, if you have received a professorship in a subject over the years,

00:14:45then you should observe these changes and also try again,

00:14:49to get the better out of it rather than demonizing it, pushing it aside and saying,

00:14:54Yes, you know, my tools, which were mentioned here were InDesign, mentioned was Figma,

00:15:00mentioned was really no idea, that can't even compare, because there is

00:15:04still the switch over there, to the right in the corner, have you ever seen it, no one knows it.

00:15:08But that is indeed insider knowledge to think, and that has been for a long time.

00:15:13I mean, especially in the digital environment, it has always been the case that things

00:15:16have changed quickly, just got to be honest, so the halftime, that

00:15:20Then I really don't get all that long, since the Internet has been there, when I think about

00:15:24my career alone, which is already a few years old, but how

00:15:28many tools have changed, which things I no longer do.

00:15:32For example, you had the Roms that I used to produce, they no longer exist.

00:15:35Back then there were tools like Photoshop, that was the non-plus-ultra in registration,

00:15:39but there was simply nothing else.

00:15:41Yes, but the speed has, of course, become cumbersome.

00:15:44And other companies are entering the scene.

00:15:46Totally, totally.

00:15:47But you could also say that no one knew that ten years ago. There are indeed things now that

00:15:52have emerged, which have now actually become standard in the screen design scene, like

00:15:57I think, for example, the Nomus Ultra tool, there are a few others as well, but

00:16:00that is actually the standard tool that is available. And of course, they will look to

00:16:05react to that. But there are also situations where these major

00:16:10market-defining players are displaced by new tools that come up. And I believe,

00:16:15the real shift marker is that Figma was a consistent development

00:16:21of Photoshop, from Fireworks, Adobe XD came next, then there was Sketch on the Mac,

00:16:27basically all these programs we had. And now we are in a situation where

00:16:33our toolkit is fundamentally changing, because theoretically it is no longer necessary to know the last

00:16:38hidden button in your software at all, because in an emergency, I mean, before,

00:16:45you would have googled, in an emergency you now ask the AI, and ideally it just

00:16:50rebuilds it for you, because the workflow for this functionality is simply automated, and you

00:16:55don't even have to worry about where it might be, allowing you to focus much more

00:16:59on this creative process and especially on the collaboration process with AIs and other

00:17:04people, basically. And I'm still not sure how the, I mean, Club Design is a

00:17:10good approach, but it will not be the end of the line, because it is still

00:17:14so heavy and pumped, and I can have a bit of influence, but it is still more

00:17:20like, yes, I'm moving iteratively through it. So the UI for future

00:17:28design and how designers will work in the future, and designers in a broad sense are not just those who can move pixels.

00:17:35But those that recognize the needs and then build a new product from that and want to design it

00:17:43This interface is still not there, well it is still very text-heavy, we have already discussed that

00:17:49We are still in the AI age in a text-based age we are only slowly transitioning into

00:17:55graphical interfaces, and in my opinion, these are necessary to achieve a

00:18:01high level of abstraction and a high speed then also for the

00:18:03Reaching people is better than doing everything just through text. We also discussed the topic

00:18:07of controllability when you have multiple text fields. Unfortunately, this has nothing to do with

00:18:13design, but still, when I tried it out over the weekend, I didn't

00:18:17really have time for it because I had another AI project, but I

00:18:19want to briefly talk about that, because it again shows a bit how it

00:18:23frees you, because you said earlier that you have

00:18:29Nowadays, you don't search Google anymore, you ask the AI. In our annual summary episode back then,

00:18:36I mentioned that I ordered a Pridde from Even Realities, which has some embedded

00:18:43surfaces so that you see a kind of green monochrome screens as a carrier from the outside.

00:18:49You can hardly see it from the outside, but you see it from the inside. And I unpacked it again and wanted to

00:18:54play around with it a bit and was a bit disappointed because they now have some sort of

00:18:57App Store, but in that App Store, there isn’t really anything decent, nothing that I

00:19:01considered decent, and I also talked to Claude and then Claude together with me

00:19:06effectively brought Claude onto the glasses. You still need a computer that you use as

00:19:12an interface, but at the end of the day, that thing built an app with me in two or three hours,

00:19:17which has a frontend on these glasses, this glasses, which uses microphones, this

00:19:22glasses shows me things, and what was the result? It was a meeting assistant,

00:19:26that would help me, as we're currently discussing, what is

00:19:30maybe your opinion on it? What questions might one pose to them now?

00:19:35How could you also respond to questions that are asked of you, and you sit there

00:19:40and think, damn it. What is that? You're wearing glasses that

00:19:45somewhat make you resistant in conversation and support your arguments, because in the background

00:19:51the big Opus model is running and someone in front thinks you have glasses on. No worries, all

00:19:55colleagues, I’m not wearing the glasses at work. You would also recognize them,

00:20:00they definitely have a different frame than my classic standard glasses, but I

00:20:04found it totally cool at that point to be using this thing as a kind of throne, what is the added value

00:20:10for a variable that you wear in front of your eyes, that could help you in your daily professional life.

00:20:16And that came out, where I then thought, I would never have pieced that together with Googling in my life

00:20:21at least not, well I probably could have, but not in the

00:20:25time and with the patience. And now I'm constantly pondering in circles, what else could I

00:20:29add. And in doing so, the other thing helped me, which brought out a Tropic

00:20:34yes, well, it likely would have worked with the previous version too, because they made

00:20:38that very small leap from 4.6 to 4.7 with Opus. And that was also

00:20:46a point where I then thought, damn it, it's just a small point, but in the

00:20:51background, they have already turned a lot, because if you look at the benchmarks

00:20:55again, the system leads again. It’s not yet there, what one would assume with

00:21:01Mythos, if that interests you, please listen to the Mythos episode again, but still

00:21:06with what speed these models are progressing, changing their working methods.

00:21:11They now have, for example, if I may briefly mention, they have different modes,

00:21:17how much effort the model makes, to put it very simply.

00:21:21Entropic has also been criticized recently because they set the mode to effort on medium.

00:21:27It is suspected that Entropic has slight data center capacity issues

00:21:33and that they have therefore set all possible models to medium in the standard, which

00:21:38needs to be changed manually in the command line.

00:21:40Basically, if our data center goes down, we might as well set everything to medium instead,

00:21:46step on the brakes a bit.

00:21:47That was also around the time when people sat there, damn it, Opus is

00:21:51no longer in 1st or 2nd place, Opus is more like 10th or 15th in coding, precisely because of this

00:21:57effort stuff.

00:21:58But you can still switch to Max manually, everything nice.

00:22:02they introduced a new mode. It's worse than Max, but better than High,

00:22:07somehow X-High or something like that, they call it, no idea. Then Opus also distributes

00:22:14work assignments to smaller models. They don't put too much effort into finding the

00:22:22solution themselves, but it's still more than if you switched to High, but less than

00:22:26if you switched to Max. You can still switch to Max, but there's a different

00:22:31problem; they recalculate the tokens, which means they divide the texts, I say it in smaller

00:22:36chunks than before, which means the tokenizer burns more tokens because it's simply much

00:22:42more detail-oriented. But the catch is that if you take that for a seventh model and

00:22:48you set everything to Maximum, you could incur up to 40 percent more costs. That's

00:22:53pretty drastic, especially if you, as a company, say, okay, I’m now going to

00:22:57use Entropic or because I want to operate it in the EU data space at Google or Amazon Bedrock and

00:23:03then suddenly you get hit with a 40 percent surcharge. You really have to think about that,

00:23:07is it really worth it to me to give maximum effort for the largest model?

00:23:13or would I rather go for Medium? Yes, yes. Yes, exciting. Exciting question.

00:23:18So that's also interesting. On one hand, you're happy about the accuracy of the

00:23:22tokenizer because it takes it extremely accurately. On the other hand, you get nothing for free

00:23:28in the world. And in this case, also tokens. And while everyone is celebrating,

00:23:33saying, look at Entropik, what a cozy little shop, he’s letting the prices

00:23:38for input, output, tokens to match, it’s naturally a simple calculation when

00:23:42the actual token burn, so the number of user tokens in relation to the

00:23:48tasks back then increases. I have a question and a comment. Of course, it's an interesting story with the topic

00:23:58of tokens in that we initially think it's good that it reflects after

00:24:05Optimum remembers from the side of Tropic, saying we see that we

00:24:12see that it doesn't always have to be the same model, so I can't say anything

00:24:15GPT, or something more; I had it in between at some point, where they also had such a

00:24:20automation in some version of GPT, where I think you also didn't know,

00:24:23exactly which GPT was actually activated in the background. They had a sort of auto-mode

00:24:28for a while, I think they are back to more sensitivity. Also the other mode. That raises my

00:24:32question because, in principle, in my experience, one or the other listener,

00:24:42I have also entered the OpenCloud world and, in principle, have a combination of

00:24:47local models and online models that I like to use for different tasks

00:24:51and I've also tried to build something that should distribute the work.

00:24:57I'm not 100% successful yet, better than I am, and still worse on the monitor

00:25:00that it actually works at all, that I have built that.

00:25:03Or also, that the devil models aren't constantly deceiving me that

00:25:06Just the thing about the billing models, I need to check my monitor a bit.

00:25:09Skills.

00:25:10So not always looking at the credit card.

00:25:12The credit card is honest.

00:25:14The useful online usage has been returned.

00:25:17These seem to be the things that I've triggered there.

00:25:21Nevertheless, it's a really exciting question, how the 4-7 model decides now

00:25:26in that case, when it acts in which moment.

00:25:30And you really need a lot of context, even as a model, you need a lot of context

00:25:34to then decide whether this market request is a complex

00:25:39coding that I need to do, or it just wants to understand how the weather is in 773 days.

00:25:44So that's how it is...

00:25:46Which could also be completed if it has to calculate it itself.

00:25:49Exactly, that would be a decision, it would have to say, no, sorry Mark, what does

00:25:53that mean?

00:25:54No idea how the weather is in 773 days.

00:25:56Yeah, that could be a quick answer before it starts programming wildly

00:25:59into something else, how could it, how will that be solved?

00:26:03So before I get back to this, maybe a clarification, because

00:26:07what I just said with this XE, there is also the second one, namely Adaptive

00:26:13Thinking Strategy, something like that. You can always turn that off too. This is the

00:26:17delegation to other models. One is the effort that the model makes, the

00:26:20other is delegation. You can control both via parameters, turn it off. I

00:26:25have it on, and when I use code and give it a task, I always tell it first,

00:26:32plan your task, implement the plan, then it tells me, yes, for the implementation,

00:26:36I recommend sub-agents, because they have a fresh context and you didn’t see that.

00:26:42And in that breath, it sometimes writes to me like, that's a small

00:26:48coding task, that's a large coding task, and then distributes it accordingly to Sonnet

00:26:53or takes care of it itself. I think it also felt a bit like this,

00:26:57Advisory Strategy explained that it says a model gives a task and controls,

00:27:02the other models are allowed to work and then swap models out of the topic,

00:27:08token consumption and costs. And I think again that's a double-edged issue. One

00:27:15is of course for the companies that perhaps want to spend less money, or for the

00:27:17private person. And the other is Entropic, I think, is happy when the data centers

00:27:22are not always running, 100 percent running. What it then ultimately bases that on,

00:27:28I don't know. I have the feeling I have sufficient transparency because it just tells me

00:27:35and shows me. Like, with CPT I wasn't so sure. I think back then they

00:27:40didn't even show which model you were currently using. I just wondered,

00:27:43why the answer was the way it was. I don't think back then they showed you,

00:27:47that it was already an answer from CPT. Here please enter the number, that was

00:27:53correct back then. Yeah, nowadays you don't even know anymore, with which versions that

00:27:56was always the case. But I found this distribution quite good at first.

00:28:02What it's based on, I don't know. Yeah, we can also say, we can also say

00:28:08we don't know something. It's good too, but this sign knowing brings me naturally to my favorite

00:28:13topic of trust. So, in this case, you trust again the underlying

00:28:19Company, because you have talked about this company a lot in the last few months. We have

00:28:25It starts like all good companies and it has principles and it has a principle.

00:28:32Greetings to Alamno.

00:28:35No, no.

00:28:37Yes, other end of the alphabet.

00:28:39Okay, okay, now you're there.

00:28:41Because I was initially like it was a must.

00:28:43That was for the intellectual listeners.

00:28:46As I have again all now please helmets to the mailbox.

00:28:52It's nice that you are excluding me.

00:28:54back to the red driving. Yes, the exciting thing is, we build trust as well,

00:29:00about branding, about keeping promises partly also, that things work well,

00:29:05whether it's bank coding or now like in design. So I'm slowly building more trust

00:29:10in the direction of Entroffic. But fundamental questions like, for example, how does

00:29:17such a model actually work, how does this model select something, remain

00:29:21hidden from us. Whether we must know this is another question, as long as you, as you just

00:29:27described, have a good feeling about it in some way, for example like a tokenizer,

00:29:32that's still completely sufficient, but it's just

00:29:36such a really exciting thing because this complexity, which may lie behind it,

00:29:41doesn't always have to be understood by everyone. That's the advantage of

00:29:45interface design, I don't have to mess around with data or anything

00:29:49else. I can have it displayed well. I don't need, I don't know, SQL statements.

00:29:54can write well, or be able to do things well, rather than just that

00:29:57it can also be done differently. That's good. It's always just a question of,

00:30:02to what extent we are allowed to pass on this layer. And how much, in principle,

00:30:08trust must be built through GUI elements, through other topics,

00:30:14that show me what is happening, as long as I really am human in the Belug

00:30:19And I think that's going to be very exciting. So this, who would now state the thesis,

00:30:24who delivers better at the moment, has a better grip on the outcome, for model providers,

00:30:30for software providers who work with the models, whether these are any services,

00:30:34that use various models in the background. There are also in the design area,

00:30:38for example Crea.ai, which offers many different image models. You don't have to pay anything,

00:30:43you only pay once, then you can use all the really models there.

00:30:46and depending on what is best, they also have an interface for that, which must be trusted.

00:30:50I believe those who are currently laying a good foundation will essentially design,

00:30:56those who will have the greatest success, regardless of whether they have the best models.

00:31:01Because I believe this trust in what I pay for, what do I actually get out of it,

00:31:07is a certain consistency. We just mentioned this with crop design,

00:31:10that if I manage to achieve consistent results through the design system in coding later on,

00:31:15or also with illustrations or posters,

00:31:20that I design, because I have established a certain design system with cloud design in this case,

00:31:23or with Google Stitch at that moment, then the results are more consistent for

00:31:30me. I have a higher trust that I will accomplish this work in five minutes

00:31:34and not have to try twenty times and such things. So this topic of trust,

00:31:38which has been on my mind all the time, will gradually become more and more the central one.

00:31:43Questions for all the topics we can create with AI. Which systems, what kind of

00:31:48Wordpress, what agentic networks will also prevail? I just thought,

00:31:53I'll do what you just said, I won't look, I'll ask the AI, then Opos told me something,

00:31:58like Claude threw something into the ether. I didn't understand it, I said,

00:32:03I simply didn't understand either. That's why I can't say it now,

00:32:06maybe we need to follow up on that in the floss check, I don't know. A second point,

00:32:10that came to my mind, so from the time language, that was also a topic from

00:32:13last week, I had given the topic of refectories and I thought, come on, this should be one hour

00:32:18over, right? It took two days. Two days. And somehow I knew,

00:32:23when will he finally be done? Yeah, I mean, or has he already somehow

00:32:27here over-engineered it, right? So the idea is, every letter a line or

00:32:31what do I know, right? Just fewer lines doesn’t automatically mean

00:32:34good, because line breaks help people tremendously in understanding texts. So

00:32:38from the one I have no idea what he is doing right now, but he's been doing that for so long

00:32:42that I think I would have been really happy if he had just told me at the beginning

00:32:45that.

00:32:46So first of all, don’t worry, your Max 20 quota is more than sufficient.

00:32:52And secondly, to be honest, today you don't need to expect anything from me, because I am

00:32:56going to be busy with this for at least the next day.

00:32:59That would have been information too, but we recorded another episode about that

00:33:04as well.

00:33:05The one I believe hasn’t even been aired yet, I think it’s coming soon,

00:33:08This is still our bonus episode for the slow cucumber season, in case we need to bridge something.

00:33:13Yes, although I'm not really sure if we need to wait out this slow cucumber season.

00:33:18It's just a bonus episode.

00:33:20At the latest, we’ll release it at the turn of the year.

00:33:23No, that's much too late, Mark. It needs to wait a bit and then see the light of the world.

00:33:28I have the feeling that the speed, as I said, the speaker keeps going over,

00:33:32the speed is increasing so rapidly,

00:33:35that I unfortunately have to say, the one or other podcast episode we recorded three

00:33:40or four weeks ago has not aged so well.

00:33:45No, everything is still good now.

00:33:47I'm fine with three, of course all episodes are good, so you can also listen to our older episodes

00:33:50listen.

00:33:51But I believe that is actually the case, right?

00:33:52I think we're currently broadcasting once a week.

00:33:57To be honest, that's almost too little given the pace at which

00:34:01things are happening out there.

00:34:02However, to be honest, not everyone wants to

00:34:07be confronted with the topic every day, and probably somewhere in between is the gold in moderation.

00:34:13I would have thought you would ask me. It’s nice that the agent has been going for so long. But what does he do at night?

00:34:20That would have been a phenomenal transition. Such a phenomenal segue. Would that fit our red thread?

00:34:27That's why I would briefly say, hey Jens. What do you think about the fact that this thing has been going on for so long? It's crazy, right?

00:34:32Yeah, I think that's crazy. So I've only ever experienced it this way, that even when I triggered my code,

00:34:38it always stops after three or four...

00:34:41Routines that have been pushed through and some magic code he wrote, which I then just post or

00:34:46have deployed somewhere, then stopped somehow.

00:34:50Actually asked what I should do or what he should do or what I want to do with it or something like that.

00:34:55Well, actually, it's like that in my environment before I, of course, wrote this article and that night you just mentioned,

00:35:01I believe you called it Night Shift, which you recently wrote about

00:35:06where you built something small that actually helps that an agent

00:35:11can also run for a longer time and not just produces nonsense over a longer period,

00:35:17because it doesn't actively ask the human in the loop first, but actually

00:35:22manages to deploy quickly when it finishes overall.

00:35:27So I have this thing called Night Shift, it's a GitHub project and I'm sure we'll link it in the show notes, I'm sure I know who writes the show notes.

00:35:37But it addressed this issue a bit, damn it.

00:35:40I'm tired or I want to leave the computer, but the thing isn’t finished.

00:35:45And back in the day, you might have waited until it asks for my feedback, and then you might have turned off the rights or something.

00:35:51And you always had this problem of how it runs in the iterations.

00:35:54That initially helped Entropic a bit with, there are different parameters,

00:35:59with which you can start Cloud Coach. Some of them are called Dangerous something,

00:36:04some are called Auto Mode. What did those things do? They always told yes and

00:36:08no to inquiries and basically made decisions independently.

00:36:12If you used these modes, especially the Dangerous one, then you had the issue,

00:36:17well, if it goes very wrong, it leaves the root folder and writes funny things somewhere

00:36:22because it thought that was useful for providing performance, because the longer

00:36:27such a context runs, the more likely it is that it forgets what I should do against my

00:36:31director as it progresses, whether to leave or not and oh, it would be

00:36:34nice to build that over there, yes, so in the sense of, oh look, a

00:36:39squirrel, I have a much better talk, I’m now watching the squirrel

00:36:41go by, just to stay with the imagery.

00:36:44Mhm.

00:36:45That means you constantly click on approvals and for longer tasks you must

00:36:51then compress the context. Claude sometimes also forgets the plan, repeats tasks, uses the

00:36:58wrong package manager. Yes, you can work with the Dangerously Skip Permission, but

00:37:05still. Yes, after 20 minutes it's over. Not because Claude is finished, but because he simply

00:37:11lost the thread. That's why I built such a combination of skill and Bash script.

00:37:17It tries to solve that and it's called Cloud Night Shift on my side, I published it as a GitHub project

00:37:22and it generates a task description and an autonomous setup.

00:37:28So you define where it has to be done, then it states in a runbook with concrete

00:37:35steps based on genre templates, refactoring, feature, migration, bugfix,

00:37:41test in cleanup, DevOps, documentation. And this runbook is processed. For that,

00:37:48so-called hooks are also defined. A hook is that, for example, after every context compression, it automatically

00:37:53says, read the runbook for the next open point.

00:37:58It doesn’t run fully. There’s also a pre-tool-use hook. That means that

00:38:02destructive commands, what a heavy word late at night, are blocked,

00:38:08before they are executed. That means you can also leave it alone because destructive

00:38:13commands like delete are marked, it tries to extract those. Additionally, the whole thing

00:38:18is also enclosed in a MacOS sandbox. It’s somewhat deprecated, but you can

00:38:25still define a profile and it prevents Clawed-of-Colonel level access to it,

00:38:30for example to unlock his project folder and leave it. And additionally there is

00:38:37also a watchdog with heartbeat check, so to speak, if something goes wrong,

00:38:42in the system, that it notices it and does not try to get stuck in endlessly hung

00:38:48processes. Each management goes through this validation of structure, quality,

00:38:56safety, and then you have no vague interpretation, but really a step-by-step description,

00:39:02of what it should do and it also adheres to it. And for me, that's really a personal

00:39:07milestone. I am currently integrating this into my other tools as well and this way I get

00:39:12to have multiple projects running in parallel. Each with its own runbook,

00:39:16each a separate sandbox, each with its own heartbeat and code becomes somehow

00:39:21from the interactive tool to an autonomous worker, who can also deliver overnight and

00:39:26cannot be left alone. And the difference is, there is no permission alone in Dangerous.

00:39:33makes Cloud autonomous, but rather this interplay between this script and the sandbox

00:39:40is precisely what is actually, how should I say, important at this point

00:39:48is. And additionally, there is also a Deemen service, meaning it works with a

00:39:55Inbox and Outbox infrastructure, which means there is basically a folder where we input the

00:40:01tasks, where the results are delivered. Every task that is submitted

00:40:06starts with its own fresh session. This way, you don't have

00:40:10such a context rot, so that if you stay in the same session constantly,

00:40:13it kind of messes everything up over time. It has a

00:40:18behavior-driven environment that checks if something new is going on with the demon or if the

00:40:26demon, i.e., Claude, is working. It has a workspace memory and decisions MD to

00:40:33share its experience across contexts. And the whole system is highly configurable.

00:40:40And that's how you can do it with the two skills. One ensures that it

00:40:45effectively self-monitors, starts itself, runs through the

00:40:51work, performs the task by compressing the context, taking the most

00:40:54important things with it, while the other operates in such a way that it effectively

00:41:00waits for tasks, no matter what tasks it has, and when you

00:41:06appropriately retrieve everything, then you have here at this point the

00:41:11option to really not let it run in the background like with Open Cloud or so,

00:41:16but rather very minimally with the

00:41:21tools and possibilities that the command line and the CLI, that is,

00:41:26Cloud's terminal interface, allow you. And especially this cross-run memory,

00:41:32yes, that decisions are retained, that loops check, that

00:41:38they are not, let's say, endlessly repeated, that when errors occur, it

00:41:43stops, that it is traceable, because everything it writes is either in the

00:41:49dedicated outbox, folders or secured via GitCom, you can really

00:41:56manage to have a digital worker who is learning in different ways,

00:42:01yes, sitting on the tasks, tracking the work, and ensuring that it effectively

00:42:09tries to reach the goal and that, if it should happen to abort in the meantime,

00:42:13like, I've tried something and I haven't succeeded for the 3rd, 4th, 5th time,

00:42:18how is it supposed to repeat that and, let's say, burn tokens or run through,

00:42:22the tokens expire and then tries to somehow continuously restart things,

00:42:26to prevent this from happening, it has also integrated appropriate protective measures and if it

00:42:29interests you, I can gladly note in the show notes where the skills are available, and then I would

00:42:35also be very happy to receive feedback from higher-ups. Yes, definitely do that,

00:42:43that is exciting. I also have a few things that I've already had in my own setup,

00:42:47you pointed out a few things to me, like this automatism and the topic of this

00:42:52multi-sessions, I hadn't had that in my installation so far, although my

00:42:56setup still revolves a lot around the creation of knowledge and the development of the second brain and

00:43:01stuff like that. I have word-finding issues with you and a little less about the coding.

00:43:09I was recently admonished by my Open-Claw instance, which actually said to me,

00:43:15Jens, weren't you actually supposed to implement projects? That was not a highlight

00:43:21by the way, that moment, because it noticed that I was spending all my time optimizing.

00:43:25I'm still training during the optimization of my walls, this knowledge worker here

00:43:30integrated into the centenizer so that it can't possibly be a dangerous source.

00:43:35Archive was built in the background, it runs locally and at night when the

00:43:39computer that doesn't get any input from me basically also slowly searches through the local

00:43:44models searches through the texts again, creating new

00:43:47connections, clearing things up, and such things

00:43:51I've implemented all of that; in the end, of course, I wanted that as well

00:43:54For that, we always tried to set up real projects with us, and that was a nice moment now,

00:43:59it's currently a time constraint that my Open Cloud installation is asking me that.

00:44:04It does that because I have of course provided it with the scripts and other themes,

00:44:07that if I always have to agree, it also has exactly this kind, which we often

00:44:11criticize here, that the AI always says yes, and that was the best idea,

00:44:16anyone ever had, whenever something is prompted or so, that mine is just set.

00:44:21mine should be that it doesn't say that all the time, and that's why it comes

00:44:28my interpretations and also such a nice moment that she then realizes

00:44:32Yeah, she had then brought Robert Hol in

00:44:35and then has been moving around the whole time and maybe one could already do something productive earlier, which is also quite exciting

00:44:41application scenario for kis actually in the future is us from such a

00:44:46procrastination is sometimes not even that but it is indeed sometimes just a thirst for knowledge

00:44:51experimenting, stepping into the famous rabbit hole in Drutter and doing something there. And just like otherwise,

00:44:57sometimes you also need a little push or the famous kick in the butt from

00:45:02someone to get back on the right track. So when OpenClaw tells me to take out the trash,

00:45:08then OpenClaw is also a whole different story. Maybe. But since you mentioned

00:45:14OpenClaw and the models earlier, I just remembered,

00:45:18sorry, I have to down the nerdy rabbit hole again. And namely,

00:45:23you just mentioned before, whether he really uses the simple models or whether he

00:45:27only suggests them. I saw an installation where Open Claw is controlled by Open Claw from

00:45:33Open Claw. So someone who has multiple Open Claws, each Open Claw has basically

00:45:39a task. So one does knowledge, the other does research, the next one does

00:45:42this that and the other. On that basis, he has limited the tools and the rights of the future and among each other

00:45:47the Claws then talk and one Claw assigns, whether that might be a bit of using a cannon

00:45:53to shoot sparrows. That's one question, but the other question I found, I

00:45:58thought it was quite funny. As in, you now have a Claw that takes care of

00:46:02the things that might also go wrong, he will be handled a bit differently

00:46:04than the Claw that picks things up, than the Claw that communicates with you over Telegram

00:46:10or whatever and grabs his other Claw brothers and

00:46:16then closes that Porz and who knows, I found that quite interesting. On the other

00:46:21On the other hand, I wondered, uh, Apple will be pleased if everyone also gets a

00:46:25digital Mac that speaks up and says, this is the Mac for the

00:46:29knowledge archiving, this is the Mac for, and this is the Mac that ensures that the

00:46:33Mac gets kicked into gear and uh, brings the machine out and prevents procrastination.

00:46:36That's an exciting question, when is this, this, this chain basically over?

00:46:40There are many things in the multi-age and topic there, that one says, okay, such a

00:46:43accusation thing. You also had the topic of board stories before. That you have such a

00:46:48advisory board. I still have that. It's actually very cool. I never said,

00:46:54that I don't think it's cool now. I just didn't know that you didn't have it. The feeling,

00:46:58you don't really care anymore that we are now also gone from this board, because you were directly

00:47:01taking care of it. Because that is in principle a little bit, thank goodness, they still

00:47:04no sense of time. That’s something nice at the moment. I often talk about that,

00:47:07that these things actually need a sense of time. Sometimes I'm quite glad when

00:47:10when you revisit such an old thread that he hasn't addressed for over 4 years,

00:47:15not for a long time now, but your advice report will also be happy when he comes back to it in 2 months

00:47:19for the first time again, then who will react as if you

00:47:23just talked to him about that a second ago, that's something

00:47:26I'm also grateful for.

00:47:27Luckily, before you know it, Navi, it can get lost ten times and it says,

00:47:33never, I said dollings, yes, the honest reboard won't say that quickly either

00:47:37nice that you are here, but now I don't feel like it anymore.

00:47:42Yeah, but look, maybe we need to incorporate that too, so maybe you have to

00:47:44what's also coming and it will also come that these things react like this

00:47:47little moment that I now have in my eyes clear, the installation had from me, that

00:47:51I just told you about, where it pointed out to me, to follow the

00:47:54projects again.

00:47:55Yes, but of course this delegation of who does which task and which

00:48:02task is even meaningful in these workflows, to do by whoever, whether that's

00:48:08a person, a local AI model, a high-class model, or a

00:48:13low-class model. This also brings me a little to this topic, where I've thought this week,

00:48:17where I said, somehow, when you look at the current situation we have in the world.

00:48:24And I think it's that nowadays alone in

00:48:29America, I believe, 50 percent of users simply throw documents into their trusted AI

00:48:35and let this AI explain these documents. In private settings, it's most likely

00:48:39some medical reports that one does not understand, X-rays, billing, other topics,

00:48:45insurance letters or something that an AI can wonderfully explain. In the work environment

00:48:49it is often also the case that I say, there is simply also the PowerPoint,

00:48:54Word document, the requirements document, that someone has written long about 50

00:49:01pages in Word, is then also taken and either through Copy-Pilot or other AIs that one then has

00:49:06available in the company, is read out. And then it's discussed what one should actually do with it.

00:49:10And then you get to the point where you say, okay, now I have to

00:49:16send my answer back to my boss, my team colleague. And because I

00:49:21know that he would like it in PowerPoint format or in a Word format,

00:49:24I let the AI generate a Word format.

00:49:26I then send it via Teams or email to my colleague.

00:49:30This colleague at work, of course, today I won’t say anything else,

00:49:33he takes it and gets it analyzed by his AI,

00:49:35generates a summary, and now we can think further along the chain,

00:49:40about what has been happening all this time, and I've thought about that this week.

00:49:42That’s also a kind of token theater,

00:49:45that could be very, very quickly ended

00:49:47and I would now claim that millions and billions of tokens,

00:49:52could be saved instantly if we just allowed the possibility, and the technology is there, we have e2a

00:50:00pipes, we have the MCP and other things that we basically have, that are already defined, if we would allow that

00:50:06in certain situations

00:50:08the AIs could just talk to each other directly, that my COBOL version, for example, can talk to your COBOL version

00:50:15And not like we do sometimes, I have to admit, you with your AI a

00:50:21summary for a possible episode, basically as a document shake, that I then

00:50:24take into my AI brain and then send you back documents and say,

00:50:29look, let’s rather discuss this right now. So I believe there is still

00:50:32this workflow, now off to the side, if we now bring ourselves into it as humans,

00:50:37who basically also belong in this workflow and decide,

00:50:42Who is going to do the next task, is it the high-quality AI or am I the one who can

00:50:47basically fulfill this task independently? Honestly, there is still a lot to shape.

00:50:51And a lot to save. So if I think about it, I create a presentation with AI, so that you

00:50:58can summarize it with AI, to then get the AI to tell me what should I respond to the market, which then

00:51:04builds a presentation that the market can again summarize with AI.

00:51:09I believe we are back to the topic we had recently.

00:51:12Maybe the way we collaborate is also changing, and I hope strongly that

00:51:16there is something more here than just connecting pieces between skills, that are between

00:51:24content creation and representation.

00:51:28I really hope so, I hope so very much, but I believe we are actually no longer

00:51:36And we will also establish new ways of working, but this is just an observation I am currently making.

00:51:44Because I often do this. Sometimes I get something sent, I don’t remember now whether this PowerPoint or this Word document has also already been generated.

00:51:51Sometimes I suspect it, but actually I take it and then get the summary of what is written there.

00:51:59And I’m no longer sure if that’s the right way of communication.

00:52:04Previously, one would always say, if there’s some email ping pong or something like that and

00:52:08things are taking a long time in a company, it’s always good to pick up the phone

00:52:11and talk to someone directly.

00:52:13That is sometimes much faster or just to schedule a meeting, where you then also

00:52:17look at a topic together and solve it together.

00:52:19Now, there’s not always the possibility to do everything synchronously.

00:52:25So we can all talk with everyone.

00:52:27Therefore, we need some kind of asynchronous communication and for this type of asynchronous communication

00:52:32we have invented all possible document types. I had on

00:52:35Hesse to do this. PowerPoint actually to present in front of larger

00:52:38groups, Word documents as text documents, for exchanging longer texts in that

00:52:42moment, email for exchanging shorter texts, tweets for exchanging

00:52:46very short messages, SMS and so on. So we have

00:52:50invented different ways and now we basically have

00:52:54AI that can help us in all sorts of ways. But we use it

00:52:59I think in a kind of way, yeah, I don't even know what the right word for it is, but

00:53:03like, yeah, like toddlers who just play around with it like a toy and actually still

00:53:09not doing it right. And just like you said that PowerPoints and emails are for certain

00:53:14target groups, podcasts are also suited for certain target groups. And before we turn to the next

00:53:21chapter, I want to conclude, I thank you very much for being here,

00:53:27Try Claude, how is the design, what is that thing?

00:53:32It's called Claude Design. Yes, yes, already.

00:53:34Thanks, look. I couldn't remember from 55 minutes and have already forgotten, it says

00:53:38not procrastination, it’s just the calcification of age.

00:53:43In this sense, Jens, thank you for being here.

00:53:45I thank you for your patience that you have shown us again.

00:53:48And you notice, it doesn’t get boring.

00:53:51Until the next episode, see you then.

00:53:53Bye.

00:53:54Bye.

00:53:57Welcome to Think Different, Think AI, the podcast by Mark and Jens.

00:54:03Two tech-loving minds that not only talk about artificial intelligence but live it.

00:54:09Here you find clear categorizations, real practical insights, and a fresh perspective on what is possible.

00:54:16Understandable, critical, and always with a wink.

00:54:20Encouraging for thought, for a chuckle, and above all for discussion.