2025 Wrapped

Transcript

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

00:00:07Two tech-enthusiast 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 view of what's possible.

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

00:00:24AI for reflection, for a smile, and above all for discussion.

00:00:35Hello everyone to a new, nice episode from us.

00:00:39I’m Jens, and I'm happy to welcome Mark back to my side today.

00:00:43We are looking forward to doing a sort of recap with you today,

00:00:49about everything that has happened in the AI field in 2025,

00:00:53but especially with a sharp look at the tool landscape.

00:00:58And I think we're going to cover everything from image generation to simple text spotting today, right, Mark?

00:01:07I found it very nice that you said this is a nice episode.

00:01:10And I briefly thought, is this our first episode with video, oh my God?

00:01:15You're not even prepared, are you?

00:01:16No, no, no.

00:01:18Yeah, but I think today is one of those episodes, today is a consideration of time, right?

00:01:22We're just before the end of the year at the time of recording.

00:01:25It's definitely worth taking a look at where a workflow or a tool might have changed, and things we might have thought were great back at the beginning of the year - I mean, are they still relevant? I don't know, that's gone from being exciting to old news if a big company sneezes, and we can work with completely new things. That's what I had in mind.

00:01:51Okay, that sounds good. Shall we start with the simple stuff? What that is

00:01:56probably the most common thing when people interact with AI tools outside,

00:02:04I think there are probably two and a half use cases among

00:02:10the general population. I would say one is this pure prompting, looking for things,

00:02:15having things created, we can take a closer look at that in a moment.

00:02:17The other is, I think, image editing or especially image generation, not editing,

00:02:20generation, I believe. And the two and a half case, I wanted to say, because it's Christmas time

00:02:26is generating songs with some kind of video crutch music AIs that are floating around.

00:02:34What's been part one? What AI has struck you the most

00:02:40when it comes to the pure ... I chat with the thing, I prompt, I

00:02:44search for something, what was your highlight this year?

00:02:46Well, I'm quite glad that my searching has mainly stayed with Google, and when I did use an

00:02:55LLM, it was indeed Perplexity. Perplexity is quite reliable when it comes to the subject of

00:03:02sources from the Internet, with various depths of settings, configurations,

00:03:08that are available there. I found that quite nice, what I found quite amusing about all these

00:03:13text prompting stories was that almost as soon as one was released,

00:03:19the next one came along, so this cycle of, okay, Mistral, I wanted to mention, but

00:03:26somehow we say Cruck and then comes Ausropik and then comes Open May Eye and then

00:03:31comes Gemini and then comes Cruck and then comes Ausropik and then comes Google and then

00:03:34comes Gemini or at this point choose abbreviations, but this constant

00:03:40carousel and since you mentioned prompting, I find it naturally also doubly and triply

00:03:46strange how even established prompts have changed over the year. So prompts,

00:03:52that I have used, maybe we will always end up with something like N8N and so, because that has

00:03:57I was quite agitated this year as well. What I used in various process steps

00:04:03with the prompts and that were nicely established because they worked well with GPT-4 or whatever

00:04:09and then 5 came out and that still worked somehow and

00:04:13then 5.1 came out and suddenly the thing thought it needed to give me bonus information

00:04:19and suddenly I had components in my texts and my content,

00:04:24where I thought, where does this come from, then you train it out of him and it's at 5.2.

00:04:28You notice, you can just recently get the new model from Open AI,

00:04:34that you already don't have to do that anymore, that you can go back with the prompt.

00:04:37So that's already quite funny and also, when you switch between the models,

00:04:41for example, when I give a prompt to Gemini, sometimes I have the feeling,

00:04:48it's prompted, and suddenly it's faster than Obmei. It's also a bit more superficial. And that

00:04:54shows again, we're not in such a 0 and 1 machine.

00:04:58Well, exciting. But if we go back to the search, because you just mentioned Google,

00:05:05it says. So actually, over the course of the year, I don't even remember when exactly

00:05:11it was, Google has gradually basically also released its own Google AI Gemini in the background,

00:05:16which also provides search results when you actually use the normal search field.

00:05:22So a quite clever move by Google, of course, a bit of a defensive stance,

00:05:26because more and more users have also switched to OpenAI, to PepsiPlexity, to get better

00:05:32search results. They expected it. So I can talk about things much more

00:05:36naturally with a chatbot, with such an LMM, than it used to feel when I

00:05:40was using the classic search field at Google.

00:05:42There were good results, but they weren't always so context-related and

00:05:47an LMM that understands me a bit is of course much stronger

00:05:50contextually and that has certainly been an incredible competition that

00:05:54Google has felt it was exposed to and they have done well.

00:05:58At first it felt a bit sluggish in the first days, weeks, and

00:06:05it wasn't quite clear where these Google AI search results were really

00:06:10coming from.

00:06:11Really strange websites were used as references, and the results were so poor.

00:06:16Since honestly, December 2025, I have more and more the feeling that it really

00:06:23has a good result.

00:06:24If I now make a request, if I really want to search and would use AI for

00:06:29searching, I still find myself interestingly using Google again.

00:06:35So a simple request.

00:06:36You know, when something comes to mind and I don’t want to first, oh now I open

00:06:39first Chatchi.pt, we just talked about this model change, now I open

00:06:42my Perplexity app or something like that, no, then I have again this, also

00:06:46a completely normal, I am now an iPhone user, this completely normal way,

00:06:49that I simply go through the search, this search basically calls me up

00:06:52Google is adjusting with the round, because my default search engine is no longer set up

00:06:56and then the Google search result comes, the KUE search result, or it fits in very,

00:07:00very many cases simply very, very well.

00:07:02And now, when you consider, now you are a classic website operator, you are obviously

00:07:07extremely pleased that the information is now even better prepared for the customer

00:07:12so that you don't even notice that there was a customer before.

00:07:16interested in it because Google is now serving it much better than maybe

00:07:21the small preview. So I noticed, of course, that it's in Google and

00:07:26took it as a bycatch, so to speak, I believe, subconsciously. But still, the

00:07:32frequency with which I, let's say, I don’t know, was on Google.com in 2023, 2024

00:07:39has decreased a lot compared to now. Let me show you a statistic.

00:07:45Yes, so for me it’s turning a bit right now because I also find it such a ... too

00:07:51Usability and navigation-wise they do it quite cleverly when you push it back into

00:07:55the Rhine house, then it’s not immediately such a long tail of results,

00:08:00that is sometimes accustomed to elsewhere, then it has a good length, it also feels

00:08:05like a good speed, so you can almost read along with your eyes

00:08:10while the result builds up, that is good timing, then there are

00:08:13some small disability tricks they do, which help you to stay on this result longer.

00:08:19at that result for longer.

00:08:20I often still have it when I use some other element in a completely normal

00:08:22check mode, then you just have the output with the prompt and it overwhelms me

00:08:26sometimes that I have to scroll back up to the beginning.

00:08:29And I have quite a lot of stuff that I have to read there.

00:08:31Let me summarize that from the machine again, because that's already

00:08:35too long for me again.

00:08:36But Google has solved this well at the moment, I’ve caught myself

00:08:42increasingly using this colorful nice button that allows you to search further with AI.

00:08:45That's something I’m using more now than

00:08:52scrolling further down in the original Google search results part, where I look at other

00:08:57things. This means, I believe, gradually it will really replace the normal Google search page,

00:09:03and because, of course you tell yourself, the website operator, I mean,

00:09:10to have it integrated into the ads by 2026. So Google will also again provide the

00:09:14opportunities for monetization and offer companies

00:09:18options to get into these search results. I’m curious to see how

00:09:25that looks exactly like the quality, how it feels. But if they

00:09:29manage to do it as well as they have with the result generation,

00:09:34with this Flowdice result generation, then one must not, one must not

00:09:39write off Google as a major player, which still wants to stand at

00:09:45this interface between us and the web.

00:09:47Regarding writing off, you’ve now made some connections to something that

00:09:53we hinted at in previous episodes. Google had been somewhat dismissed as the one with Open

00:09:58AI recordings. Then came Bard, then

00:10:03there were some strange attempts, and now one must say that what Google has shown in 2025

00:10:11with the integration, we will also get to image, video, Nordburg LM, the

00:10:18whole, also all the tools in Google Labs, yes, learn in your style, this thing with which you

00:10:25currently, I believe, only outside of Europe, I always have to think because sometimes I have to

00:10:29additionally start a Frappe, where you can, for example, generate learning materials

00:10:34for a topic and these learning materials are prepared in such a way that they match your

00:10:39hobby, for example, or your favorite movies or something,

00:10:45so that it is explained to you in a more easily understandable way.

00:10:47And those are powerful things that Google is launching into the market

00:10:56Prest, you can already tell that Google, firstly, is a major player that has kind of gotten back

00:11:03up after a brief downturn, with its own hardware, which

00:11:10builds AI chips, its own software, the platforms, whether it's in their Google Cloud,

00:11:16or on phones as well, where they then execute the models, where they

00:11:20now also apparently sell a model to Apple from Gemini, so that they can their

00:11:25Finally get the series smart. Get smart.

00:11:29Apologies to all the Apple listeners out there. Shout out to you.

00:11:34It's crazy how Google is shaking up the market, and it's becoming clear to you

00:11:40that the money that the other players somehow

00:11:43have to collect with difficulty. Google has something with the quarterly results.

00:11:49Yes, definitely. I think that's where the power lies. I don't even know if it's truly so...

00:11:54The question is always, what have we seen, what's the tip of the iceberg, the failed attempts,

00:11:59Google has always been known for trying many things due to its market power

00:12:04and also for knowing to try many things and then throwing many on the graveyard.

00:12:08But I think it's also not entirely right when you're in such a fast-moving world,

00:12:13you also have to try multiple things.

00:12:16only those who have the necessary funds, and Google simply has that available.

00:12:20you have to do such things. And why not? I believe they were there

00:12:22maybe a little surprised, but reacted really well in my opinion, and

00:12:28since you mentioned Apple, let me say a few words about that. This is a very

00:12:31interesting topic again from a user perspective. I still find it quite strange

00:12:35especially how the connection works. Behind my series, I have the

00:12:38ChatGPT Pro version set up. When you make a request there,

00:12:43the response time feels too long to me. That's odd. It's probably not really

00:12:48much longer than if I had maybe at times just directly entered the chat prompt in ChatGPT

00:12:52or something like that. But it just takes some time in that moment. You realize that it's

00:12:55a connection first. Siri needs a moment for the speech processing. You have to pass it

00:12:59over. That takes a moment. You notice all that. It still doesn't feel like a

00:13:02really good user flow. It takes me a tad too long. And then you still get

00:13:07this reminder each time at the end that ChatGPT might make mistakes.

00:13:13So that this information, if important, is provided in such a way that they know

00:13:17Yes, you can tell me that once at the beginning, but you don't have to repeat it every time.

00:13:23That totally disrupts the flow. It makes it 0.0 fun. Where I also say, yeah, it's still a bit

00:13:31So, voice input is sometimes good, but still a bit bumpy.

00:13:38Even when I use the Pro version of JetGPT, with the voice variant,

00:13:44sometimes it feels quite good, sometimes it's a bit bumpy.

00:13:47I don't know how it is for you.

00:13:50Amper.

00:13:51The integration into Siri feels like the bumpiest thing you could imagine in this context

00:13:57can.

00:13:58it's perfectly fine if they have now integrated Gemini as their own model,

00:14:03whether they run it in their data centers or on the end devices.

00:14:09That can only be good for the whole thing, and whether Apple will eventually release something of its own or

00:14:14not. I mean, we have seen this a few times. The Apple Maps were initially

00:14:19like, how was it, streets that went right through the airport, and now it's

00:14:25It has really improved quite a bit. You just have to wait and see how it develops.

00:14:31I'm just glad when Siri becomes a bit smarter, because that is noticeable too. Also

00:14:35when Siri keeps going off, and I'm really afraid that Siri will start responding in my environment now and

00:14:40executing commands, so I need to take a little break afterwards, just so that thing

00:14:46doesn't suddenly start, and you have used it for home automation. Okay, it always does that.

00:14:52still. Yes, for me, I can control radiators, lights, doors, and the TV, and no

00:14:56idea what else I can control. But it’s still totally lame when you think about what

00:15:02is possible with voice interaction, chat interaction with other manufacturers.

00:15:09That’s true, that’s true. Shall we now take a look at, well, let’s not just

00:15:12write off. So they also have their, one or another money-saving feature like

00:15:17I believe if we look at another major player, then maybe 25

00:15:26hasn't performed as well as one might have caught up, let's say,

00:15:29someone who has been a bit reserved and from whom I haven't heard anything

00:15:32more for a long time.

00:15:33You really need to do some research.

00:15:34I haven't researched there at all, it's just stored in my gut.

00:15:37Some experiences.

00:15:38Amazon.

00:15:39Oh, I thought there would be commutators.

00:15:41No, no, no.

00:15:42Commutators.

00:15:43Exactly.

00:15:44The next C64 bot that I've been waiting for a long time.

00:15:46Bittier is re-released, right? Yes, the C64 is being re-released, but okay, we're not in a Redkasten here.

00:15:52Yes, yes, I lived through that.

00:15:54Um, but back to the topic, so Amazon with its Alexa, which I probably use the most for voice interaction.

00:16:06Rightfully so.

00:16:07Because I would say, from my perspective, I definitely have 10 to 15 interactions a day,

00:16:12that I perform.

00:16:13So, those are simple things like playing, I don't know, Germany on Knobar or

00:16:18what's the weather or such things.

00:16:20So these simple prompts via voice.

00:16:22Just called out into the room.

00:16:24One in the shower hanging up.

00:16:25But there's a bathroom hanging up.

00:16:26The other in the kitchen.

00:16:27Two different names.

00:16:28You always have to remember how you're speaking so the others don't start responding.

00:16:32But that's doable, nothing more can be done.

00:16:36But still not great in understanding. You can sometimes say which music genre

00:16:40do you want to listen to? That still works okay. But it's still light-years away

00:16:45from being a real ChatGPT, Gemini, Claude, experience in a dialogue with a

00:16:54AI. So it still feels like, okay,

00:16:58my speech is translated into text and then given to this machine and

00:17:03then this machine tries to somewhat accurately perform intent recognition

00:17:07to execute a real system within its narrow scope it has been given and somehow recognize,

00:17:13what does this user want from me.

00:17:15That's 0.0 of this AI interaction that we both have been talking about all the time, which we now have,

00:17:20the ChatGPT moment we got used to as humans two and a half years ago.

00:17:24Emerson hasn't closed that gap yet.

00:17:29And I don't know if they will, so I haven't heard anything about it.

00:17:32So I wouldn't expect anything, and I can't say anything about it, because who would be surprised?

00:17:37At least not the ones who know me probably wouldn't be surprised about it.

00:17:41I don't have an Emerson speaker at home.

00:17:45Even though some of the speakers were partly purchased by Emerson, they are still called HomePods.

00:17:50Yes, so on that front, there is at least one in every room, usually two, the house is well-equipped.

00:17:58And that would be completely enough if the HomePods got better at some point. But you can report back when yours does.

00:18:07Yes, that would be appropriate. I'm curious, I wouldn't be surprised at what comes, so I must be a big lanne.

00:18:14You have market access. I mean, that's what we've just seen with Google. So yes, they see that there's a significant share of search queries going away from my.

00:18:25search mask, going in the other direction towards AI, and they are catching up relatively quickly,

00:18:30of course now exploiting their market power. So if this topic continues to push forward with this

00:18:36speed, it will get really tough for the others. So not right now in

00:18:41in all the countries of the world, Google is the most popular search engine, but in many places around the world,

00:18:46well, it won’t be that easy to compete with this player once it really takes off and

00:18:50it already looks like that.

00:18:54That really seems to be the case.

00:18:57But on the other hand, if you consider,

00:19:00relevant standards for connectivity are emerging.

00:19:04So all this stuff with MCP and A2A,

00:19:07all this stuff about what these agentic workflows look like,

00:19:12which I always like to build with N8N,

00:19:15I can already imagine,

00:19:17that it’s not too late for the party yet.

00:19:22Because these topics, I’d say, the intelligence, the tuning, the LLM model, they can

00:19:28shop that somewhere if necessary, while all around them the whole thing with MCP and

00:19:34co. is emerging, learning, yes, MCP has something, the leading version of the protocol

00:19:40just got from Osropic, which means there is still a lot of music in it,

00:19:45this is all still new.

00:19:47From that side, I can also very well imagine that this magic, when a voice interaction

00:19:53works with your speakers, because the boxes suddenly have more RAM, more

00:19:58CPU, have local models, and then the standards are there, and you manage,

00:20:03yes, lots of if, if, if situations, that you can connect automation on the Internet with

00:20:09yourself, then I do believe that this magical moment

00:20:15from the manufacturers, who already have the devices at home with people, will then their

00:20:21Big ones, whose great renaissance is coming.

00:20:24Yes, yes, maybe, maybe.

00:20:26Let's shed light on another topic in these areas a bit, which I

00:20:31have also observed as an attempt that was then scaled back, where

00:20:37it was tried differently and where the various sized Elemambita behave a bit differently,

00:20:42that's the topic.

00:20:44At some point, the climate reasoning came up, for example, you know that

00:20:51we were involved, whether it was back at Diebzig at the time, then also with Namanos,

00:20:57when Jezipiti came up, that the models showed their reasoning so to speak, that

00:21:00they not only reacted to my prompt but thought about it first,

00:21:05what does the user actually mean with that, these thoughts also made it clear

00:21:10what they displayed.

00:21:11Yes.

00:21:12Then it was also displayed somewhat hidden again, then there are always little

00:21:16tests on how that appears from the usability perspective for the user, that one

00:21:21can expand or collapse it. Then it ended this way in a Chatchi PT5 1 version,

00:21:27I mean, where it was then automatic. Now I can select it again. Now I can

00:21:32choose again, I want the instant thinking model or I want the model

00:21:37that thinks about it first. So there’s always such a change in stand.

00:21:40Now don't hold me to the versions, lest I got them mixed up.

00:21:44But this year it was a lot of experimentation. What should we do best with the user?

00:21:50Should I, as a model, decide independently whether to conduct a deep research

00:21:56analysis? Whether I should answer very quickly? Or should I give that back to the user again?

00:22:04in hand, because in case of doubt, he wants to know better or feel like,

00:22:10being able to decide whether millions of tokens are being burned in the background

00:22:14or whether just a quick answer comes out of the model.

00:22:16So that’s something I’ve observed this year, where I think

00:22:19there isn’t yet a final answer.

00:22:22From my usability perspective, I would always say, yes, let the user not

00:22:26decide this.

00:22:27It shouldn’t work like that; the models should make the

00:22:30right choice when they do.

00:22:33It is naturally annoying when you're using cannons to shoot at sparrows, that doesn’t have to

00:22:40always be the case.

00:22:41What you mentioned about biosubility, I don’t know if you’ve seen it, probably,

00:22:48I’ll mention it anyway, I don't know when Google introduced this, yes, after

00:22:52the launch of the last Gemini model, that when you get an answer,

00:22:57you can then say below the answer, check

00:23:02your answer again, where I then think there are sometimes buttons that are hidden in context menus,

00:23:09where you wonder, don't you want the users to use this

00:23:13or was it again the nerd thinking, ah yes, I need three more buttons, that’s

00:23:18totally obvious at that point, nothing against nerds, I belong to that faction myself.

00:23:23There is still a lot changing, it is also handled differently.

00:23:28Just because someone can operate a Gemini doesn’t automatically mean they can operate an OpenAI.

00:23:34You’re making CustomGPTs or Dams here, then Manus, making some kind of knowledge storage out of

00:23:41Robi, coming with skills, now OpenAI is also bringing something with skills and then

00:23:45you sit there thinking, wow, if you can do one, can you do all, you can’t

00:23:49really say.

00:23:50You have to especially read out the specializations here.

00:23:54Exactly, you have to look a bit, but that’s how it is.

00:23:56When you look outside, Cloud might be better for coding, then it's OpenAI again,

00:24:02then you should take that one, then you should take the local version.

00:24:05However, as a user, you’re still a bit in debt.

00:24:14Where I say, I also have to take care of myself a bit about what might work best.

00:24:18It still hasn’t really clarified itself. Therefore, Google’s move is currently interesting.

00:24:21It’s a special use case.

00:24:23I wouldn’t yet think about programming an application in the Google search window.

00:24:27But for searching, that works again.

00:24:29They are simply exploiting that we are used to it and are also delivering excellent results again.

00:24:36This is a very small use case, where a clear result comes from a clear, simple request,

00:24:47where I don’t have to think about whether this should be a love research,

00:24:51also programming something. It should spit out the result in table form, but I have to

00:24:54first give it a format in world structure for the search result to be particularly good.

00:25:00Therefore, I think it’s a really exciting thing happening with Google right now.

00:25:02As for the other topics, yes, they’re still testing what the right

00:25:09interaction patterns are. I also think this, I believe I saw that with

00:25:14JCPT more often, where you then also get two results compared with each other,

00:25:18where you then say okay, which do you prefer, where I also say, there comes again

00:25:24put this thing in, then you have on one screen, on the mobile device it's a disaster, +**[00:25:29]** but if you look at a desktop screen, you still have the thing that +**[00:25:33]** I say, then I have such a long line of text, two next to each other, +**[00:25:39]** you somehow have to read and compare these two. +**[00:25:42]** It feels best to read, compare and click, I just wanted an answer +**[00:25:45]** damn it. +**[00:25:46]** I just wanted to have a directions and then you always feel like, you know my +**[00:25:49]** biggest problem is, I always find something good in both, you know? +**[00:25:54]** And then I’m always thinking, what am I supposed to do now actually? +**[00:25:57]** I really like this one part... +**[00:25:58]** Window closed, bang!

00:25:29but if you look at a desktop screen, you still have the thing where

00:25:33I'd say, then I've got this ridiculously long stretch of text, twice side by side,

00:25:39and you somehow have to read and compare both of them.

00:25:42It feels best to read, compare and click as well, I just wanted an answer

00:25:45for heaven's sake.

00:25:46I just wanted one answer, and then you always feel, you know, my

00:25:49biggest problem with it is that I always like something about both of them, you know?

00:25:54And then I keep wondering the whole time, what am I actually supposed to do now?

00:25:57I do like this one part...

00:25:58Window closed, crack!

00:25:59Do I have to now...

00:26:00No exactly, I'll copy this one part and say, this first part

00:26:05I liked that, but otherwise I would have found the other option good and

00:26:08then I just click on one and then I'm sure whether that overall

00:26:11was the better one, then it's more of a frustrating look for me, because then through

00:26:14this, there I might be overwhelmed by giving too much feedback in that moment.

00:26:20We also had a bit of influence in the introduction of our episode,

00:26:26to steer towards saying, well, what might have been really cool back then

00:26:31Did it feel like shit? What has changed in the way we handle it by now? I know

00:26:36still how happy I was about Gamma. Gamma was or is the opportunity

00:26:41to say, make new ones from texts, from existing presentations, from any ideas,

00:26:49Slides, you could create slides in Gamma on the website of Gamma, edit them,

00:26:55and download them as PowerPoint. And somehow it all changed for me overnight,

00:27:01because of Nano-Banana. Everyone who thinks, now the

00:27:07Minions are coming around the corner at Nanna. Yes, I can't really see it from that side. Please

00:27:12forgive me, it's the youthful rascal in me.

00:27:16Yes, it was great. Thank you, thank you. I'm working with Warner Bros. tomorrow, or I don't know,

00:27:22it doesn't matter. And then Nano-Banana came around the corner and suddenly

00:27:27not only painted beautiful pictures but also now with

00:27:31Gemini3, Nano-Banana has seemingly made its way everywhere. You can

00:27:36create slides. You can make infographics. If you go through Google Studio,

00:27:42you can even create 4K resolution infographics with high detail, fidelity, and with sharp text

00:27:51and really loud text. Image generation with text has always been a topic until now.

00:27:55And you can do really cool things. So my son plays volleyball,

00:28:00and then I looked at, okay, what does the photo look like when you take the photo,

00:28:03that I took of my son at the volleyball game, and Nano-Banana says,

00:28:08make a throne shot. Or you take a floor plan of the house and say,

00:28:12what does the Astrid model look like? Or you take, I don’t know, a complex topic and say,

00:28:18let’s make an infographic out of this, and then you actually get an explanatory

00:28:22infographic. And before I get into a few Google tools, maybe a little

00:28:28tip for everyone who says, Zimmermann is now talking about Nano-Banana and one can

00:28:34make slides with it. That may all be true, but then they only land in Google Sheets and I

00:28:41don’t have PowerPoint and no idea what. At this point a small tip on the side. Nano-Banana is

00:28:46also in Manus and there you can generate slides as PowerPoint. Good,

00:28:52but we were talking about Google, I'm back to Retefluss. I love the Nano-Banana stuff,

00:28:57because I absolutely have to tell you about a hypothetical, someone who would see the video,

00:29:02that we are not recording, would see me making air quotes,

00:29:06a hypothetical case I experienced in a school context once,

00:29:10where then more, let's say, someone anonymously told me, someone informed me,

00:29:16you have to do it like this. Then you somehow have notes for a subject like

00:29:20history, handwritten notes in a notebook, you have a book, you photograph

00:29:26all that stuff in a circle, throw it into Notebook LM, tell it to make me a

00:29:32infographic for the exam-relevant topics, make me a video with explanations for each exam-relevant topic,

00:29:39and you get into that learning material so quickly, you can get quiz questions,

00:29:47you can get study flashcards, you can, as I said, get explanatory videos,

00:29:52And if you have to report on a topic, you can also say, let's just create a few slides.

00:29:57Then you get slides, then throw the slides back into notebook.lm and say, let's make a video out of the slides.

00:30:02Then you get the voice-over that you can read or, of course, quote freely.

00:30:06You get such powerful tools, so I think I would have been glad to have had this as a student.

00:30:12That wouldn't be, and I think I will take something away briefly, we're also going to make another episode

00:30:23where we will take a look into the year 2026,

00:30:27it was already a little announcement because you just said that and I listened to you,

00:30:32and I find that amazing and good and you are absolutely right, that's

00:30:36really become very good, what Google offers, what the image generation

00:30:39under our co is about. Let me just take this outlook forward briefly. But of course,

00:30:45it is still annoying to say, make me this and that from the slides, and make

00:30:50me this and that from the video. We are still in such a prompt engineering phase,

00:30:57where you already have to imagine what do I actually need? That's why

00:31:02I have to prompt it so well. And we are not so much in this rather optimal

00:31:07output, almost just delegating, that I say I want to become smarter and then everything happens.

00:31:12You know, in that sense, there's still a day missing for it to be 2026,

00:31:18I believe that year will be the one where I will slowly start shifting such things, where we will

00:31:21talk less about optimal prompting in the normal application case, so for the

00:31:27private home use, and we will rather move towards saying, okay, I want a

00:31:33specific goal and I delegate this goal.

00:31:35So I completely agree with you, as it has sounded, from Gama

00:31:41to Nano Banana, from Nano Banana to Manus, so that we get PowerPoints, that's already

00:31:47a bit of knowledge, knowledge, interest, curiosity, also the right

00:31:54subscriptions at the right time, like which subscription I signed up for, so that I have which

00:31:58function available, that's where a lot of chaff separates from the wheat and at that point,

00:32:03I think we are really still in such a phase of storm and stress, with regard to all these tools,

00:32:07and of course also in finding our personal spleens accordingly.

00:32:13Yes, that's true and it's just still a bit, it's still not what I

00:32:19would have expected from 2025, honestly, because we actually, I believe, have fallen back

00:32:27a bit into the old IT view. You have to have some knowledge,

00:32:33you have to be good at it, you have to know which tool is there, you have to be able to link tools.

00:32:38You're also a big fan, and I am too, but you use it more than I do from N8N, where I can set up workflows

00:32:44by linking multiple tools together. I can do that, of course, now too.

00:32:49And that is also a huge achievement, which already allows direct workflow prompting, if I want to.

00:32:54Yes, that it can almost operate independently at times, I no longer have to bring together the individual nodes

00:33:00there.

00:33:01But it is still something that I would say is not really

00:33:07for everyday use when I just want to grab something quickly, but I

00:33:11do have to engage with the topic, to be honest.

00:33:13But even then, when we started our podcast, I had the honor to present

00:33:21which tools we work with, so that we have the sound, so that it gets deployed, so that

00:33:27appropriate notifications also come to me, for example, when a new episode is released.

00:33:33A lot has changed since then. So for those who don’t know, Jens and

00:33:39I record here at Riverside. Riverside has an API, I can link it to

00:33:44the end. This means I can get our audio tracks with the workflow.

00:33:48I mentioned in the episode that I'm starting with Ausropik and the sound jump is still being refined a bit.

00:33:55So this heavy breathing, maybe artistic pauses, because neither of us can think of anything, which is very rare, comes out.

00:34:06And then I push it to Potigy afterwards. And everything used to be done by hand, but now that I have an app not only on the Riffer side, but also in Ausropik and in Potigy,

00:34:15I essentially generate everything with my workflow after the recording is finished at the push of a button, and soon maybe even the whole topic of the cover.

00:34:27That's still a bit in the storm and stress phase, so you can basically say, okay, we have the box, I press the button and off we go.

00:34:35And in Polygy itself, perhaps one has noticed that we now have chapter markers

00:34:42or that we have more detailed show notes.

00:34:45And all of this happens because Polygy has the corresponding AI support to

00:34:50indicate where the old white men are talking about, what chapter markers they make

00:34:55under certain circumstances, and adds them in?

00:34:58And it's quite amazing how, I would say, qualitatively good content can be produced for two people like us, who mostly meet on the same day we release it, so there's not much time between recording and release, that then, let's say, 90 percent of the fat, I would say, is automated.

00:35:24Why 90 percent? To be honest, we had other guests visiting too.

00:35:30I do take a look at how it is. If both of us get rained on now,

00:35:34don’t hold it against me, but we are grown, old enough, we can live with it.

00:35:37And that is really fascinating and before I hand the floor back to you,

00:35:41I wanted to mention, I can't say a chapter without mentioning Manus.

00:35:44Even Manus has released an API by now, that you can practically, not practically,

00:35:50that you can say in your workflows, in NRN, and now please start 200 agents on Manus,

00:35:57using them for something, websites, searches, writing, generating, I have no idea what it was and

00:36:03then bring the result back into my workflow. That is really, really,

00:36:09really crazy, but also partly overwhelming. If something doesn't work, then you stand

00:36:16there in front of your declining workflow and think to yourself the world could be so simple and then it is

00:36:21usually only a very small thing that has somehow changed but it is still extremely

00:36:25powerful what we can automate these days, that’s true. I just wanted to briefly

00:36:30mention something but it doesn't take much time, and both of us then record podcasts

00:36:37and it goes so easily, of course that’s also because we are both so full

00:36:40professionals, so apparently natural talents and accordingly it is also super easy for the

00:36:44AI to take out all the airs and ours and ames and quickly

00:36:49tidy us up and make a great podcast episode out of us, because otherwise it would

00:36:54probably take much, much longer, so we can praise ourselves for that. We do

00:36:57that clearly. We could actually talk about Alkfalt, honestly, when I think about it.

00:37:00We have already done that, we already have an episode as if one knows a

00:37:02stage where we should introduce ourselves, let’s do a live episode.

00:37:06Let’s do a live episode with a real audience. That is what one says

00:37:10for 2026. We are doing a live episode with a real audience again.

00:37:14But with a foreign audience.

00:37:16We had one before with a real audience,

00:37:18but that one was behind closed doors.

00:37:20Right, it was internal, we couldn't do it like that.

00:37:22But I thought it was a lot of fun as well.

00:37:24I would like to do that again.

00:37:26Manus, speaking of Manus.

00:37:28So we also talked about tools,

00:37:32last year.

00:37:34It’s about websites, government, other stuff.

00:37:38Lava-Bill and whatever they're called. There are Replayt and I haven't seen them all as tools that allow you

00:37:45to build whole websites, applications, always something different.

00:37:48That fascinated me for a while.

00:37:51I also took a look at those things.

00:37:53That, of course, also comes a bit from necessity because I really don't need it for my application errors right now.

00:37:59I have laid down a bit of Ad-Actor, but also because I have built fully functional applications with a couple of proms at Manus

00:38:05at Manus

00:38:07that covered certain use cases of mine,

00:38:16that I thought, man, I really don’t want to deal with this topic anymore,

00:38:22what kind of files in what structure a website needs to be stored, something else, if the AI is really already doing that,

00:38:31really almost in quotation marks

00:38:34you can make a Heine. Of course, I haven’t done a security check on this thing and everything else.

00:38:38I know the whole vibe-coding community out there and the fine folks of the vibe-coding community out there and

00:38:44those who find AI coding completely awful will all say that it will take a while until it really

00:38:51comes and works a bit, and you can break in everywhere and haven’t seen anything of that and

00:38:54the security isn’t good and the code isn’t clean, and if you want to change something then

00:38:57they always write that code from scratch completely and you haven’t seen anything.

00:39:02But if I can express a certain idea as a user and have a functional prototype

00:39:11a few minutes later that works wonderfully with a login mask

00:39:18via my Googlecom and with a database in the background with the wild upload that

00:39:23then works with another AI that has been requested, then

00:39:27that is really very strong. So I would say that by 2025,

00:39:32we didn’t believe at the beginning of the year that one could enter a future

00:39:36where you can really create arbitrary applications with simple and prompt

00:39:42statements. By 2025, it was already the case that the lie was punished and showed,

00:39:47this works. By 2025, we were already able to build really small good applications that

00:39:53weren’t there before and that delivered small solutions.

00:39:56So it was pretty cool. So I would actually like to throw something

00:40:00very pragmatic into the rings, so that we can be concrete about what can be

00:40:04built, for example, and thus, as I said, it’s nothing about these workflows,

00:40:09but really, you sit there, whether in Gemini with Canvas or in Manus and say,

00:40:14you know, I need something, and the two applications that I’ve built, I still use

00:40:19very regularly. One is, in my normal professional life, I do

00:40:25a lot with the whole topic of mobile development and am also dealing with device management.

00:40:30That might sound a bit dry right now. If you take a look at Apple,

00:40:34there are many documents, but they are somehow updated relatively late.

00:40:38Beta software is never included, but there is a Gitar project from Apple. It always contains

00:40:43the new possibilities in device management for beta versions. And that

00:40:50is a Gitar project, which always consists of so-called Jammel files. I

00:40:54not to bore anyone, for the various, it basically states what I can configure,

00:40:59what that means, from which operating systems, on which devices that works?

00:41:02And now you can take the part and download it and do stuff, or you go

00:41:06and say, hey look, this works from here, it's accessible.

00:41:10It briefly describes how it is structured, so that he doesn't have to completely

00:41:14reinvent the world.

00:41:15You say what you intend to do.

00:41:16And now I have an interactive application that basically always shows me what

00:41:20new has happened. I can compare operating systems, I can compare devices

00:41:24with each other, I can compare any things with each other and even

00:41:28export the corresponding configurations, so that I am able to do so,

00:41:35I was saying, now I will try this out on a meter device. I found that totally cute,

00:41:39that looks good, it's easy to use. Although maybe one or the other

00:41:43minute was spent on the topic of prompting for this application,

00:41:47but it was there. And the second thing, we've already said it, for those who haven't heard yet,

00:41:53I like N8N. And with N8N you're facing the topic, as I just mentioned, you now have the

00:41:58amazing APIs from Othropic, from Riverside, from Manos, from whoever else. I like 11 Labs

00:42:03also, 11 Labs. And then you stand there and think, okay, there are nice APIs, but I

00:42:10sometimes just really don't feel like going through the API documentation. I

00:42:15put together a small vibe-coded app and said, you fit in well

00:42:21enough.

00:42:22Either I have the API documentation as a link as a file and I have no idea.

00:42:27I want you to take my task, then like I said, if we have the API documentation

00:42:33take it as a link as a file or as a research task and then you build me the corresponding

00:42:39JSON entry that I can then copy into the night.

00:42:42And so I built something where you just throw in, I want Potigy and the Stasienis and then it basically creates notes for you that you can copy for later thinking.

00:42:53Is it perfect? No! But it definitely helps me to take a first shot, so that I have something to work with, without having to search for it myself, think about it, do it, no idea what.

00:43:07Those are two things that I actually can no longer think about.

00:43:12Yes, and those are very good examples. I think that's still a bit

00:43:16another vibe-coding, perhaps for the listeners who aren't familiar with it.

00:43:19It really is about me saying that I program something purely through prompting,

00:43:25whether it's an app, a website, or an application.

00:43:28That's the topic of vibe-coding. So I don't necessarily need to have traditional coding knowledge.

00:43:33I don't need to know what the market is currently saying, that there’s a GitHub somewhere where I can deploy things, publish them, that there are certain file structures, that there are different file formats, configuration files, that there are programming languages in various forms at all.

00:43:49I don't need to know any of that because the machine takes care of it completely.

00:43:54The LMM simply takes my wish, translates it into code, and tries to build a functional app.

00:44:01If I'm a really hardcore vibe-coder, then I would keep just

00:44:05text and prompt with that machine all the time to get the code in the end to a point

00:44:09where I have the more refined version of what I want.

00:44:13But as Mark described, you can also use it really well as an intermediate step.

00:44:17to take, so as to possibly get started with it or also the LMM

00:44:24to say, certain parts basically only to prompt or if it is already really prompting

00:44:28parts and the LMM code then certain parts, certain conference files in the structure that

00:44:32I have already proceeded with.

00:44:33But as I said, a lot has really happened there, there are these editor providers

00:44:38out there, who aim to make it just as easy,

00:44:42as these wiki editors we once knew in development, where I then

00:44:46also pulled together UI elements with drag & drop in this niche

00:44:49go then more like providers like LevelBoneCo who are also trying to present themselves

00:44:53specifically as website providers, as app providers,

00:44:57and also utilize their models to help lawyers a bit, very,

00:45:03to achieve good results very quickly, which are also runnable, safe,

00:45:08that you can extract.

00:45:09So there is still a lot possible in that market, what I just wanted to briefly say was,

00:45:15it really surprised me how far it has come with a simple prompt.

00:45:21I just wanted a simple application and that is what came out of it.

00:45:24I'll see if I can show them to the public next year.

00:45:28I still need to tweak them a bit.

00:45:29When I said again, over Christmas, when I tweak the app, it didn't turn out cool.

00:45:33Yes, exactly.

00:45:34So next year, we're going on stage publicly and you'll release a piece of software.

00:45:39That's totally great.

00:45:40Jens, before we talk more about 2026, you already mentioned it.

00:45:44There will be an episode where we want to talk about 2026.

00:45:47want to. I would say let's make a nice gift wrap around our episode today.

00:45:52A nice wrapping paper for our episode. Some of you listening here

00:45:59and thinking damn, this is so great. The tips we are hearing here, the conversations,

00:46:04that we get to listen to, we want to make available to others. Then

00:46:08feel free to tell them that our podcast exists. Leave us a like, leave us

00:46:13a comment. Subscribe to us in a podcast player of your choice and with that, I would say Jens,

00:46:20on on. I'm listening to you, Nikolaus. Yes, it could be that we still have some things to do in a festive

00:46:25mood and I would say we wrap it up today.

00:46:29In this sense. Ho, ho, ho, ho, ho. And lastly, noted, stop, a question

00:46:38I was still open to that, I'll mention it at the very end. You asked me which

00:46:41song generator we should use. I would recommend to the listeners to listen to the last episode

00:46:47again, because at the end we included a few bonus tracks that

00:46:52we presented as music generators from MusicGPT, Sonos, or Eleven Labs with the title 'make me festive

00:47:00music on AI topics.' So you won't have to fuss next time,

00:47:07it's still, I believe, the speaker manufacturer. You stole my fluff.

00:47:15But you're right. Yes, I am right. So I won’t just keep that for you.

00:47:20In that sense, let's really put a mark on it now. In that sense, have a nice evening,

00:47:24nice day, wherever you are, whenever you're listening to us. Ciao.

00:47:27Ciao.

00:47:28Welcome to Think Different, Think AI, the podcast by Mark and Jens.

00:47:35Two technology-loving minds who not only talk about artificial intelligence but live it.

00:47:41Here there are clear classifications, real practical insights, and a fresh perspective on what is possible.

00:47:48Understandable, critical, and always with a wink.

00:47:52For reflection, for a smile, and above all for discussion.