Zwischen Bots, Agenten und 20 Kilo Fleisch

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 will find clear classifications, real practical insights, and a fresh look at what is possible.

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

00:00:24Thought-provoking, amusing, and above all, engaging.

00:00:33A warm welcome to our new episode of Think Different, Think AI.

00:00:37This will one day come out of my lips a bit smoother.

00:00:40Thank you, I'm already being laughed at by my colleague here.

00:00:43We're super excited that our last episode was so well received by you.

00:00:48Sophie, can I tell you? Jens, you haven't seen the statistics yet.

00:00:52Not that I, let's say, belong to the biggest podcasters in the world,

00:00:56But all the podcasts I've started so far,

00:01:00have not managed to become nearly triple-digit in the first week.

00:01:03I found that nice.

00:01:04I found that nice considering we were just telling a little in our small babble,

00:01:09that we are doing the podcast.

00:01:10From this side, I'm looking forward to today's episode.

00:01:13And perhaps you should be excited.

00:01:15And you all too.

00:01:16I think I can explain today what 20 kilos of meat has to do with curry.

00:01:23Okay, that sounds exciting.

00:01:25There is a lot of room for interpretations.

00:01:29I'm really curious.

00:01:31Especially now, of course, during barbecue season.

00:01:33I hope the weather is decent for you guys.

00:01:35Could this become exciting information?

00:01:37I hope it goes in that direction.

00:01:39Yes, it's going in that direction.

00:01:41And the weather wasn't quite so nice over the weekend.

00:01:43But okay, let's take a look.

00:01:45We now had an episode and as it goes.

00:01:47An episode is in the can and it turns out,

00:01:49that there are a few things that might need correction.

00:01:53need. So I would suggest we do something. I heard this in another podcast

00:01:57once. Yes, one can definitely imitate good things. We're doing a fuzz check. We will find the fuzz.

00:02:03We will find the fuzz and explain it. I call it fuzzy, you know, like fuzzy?

00:02:09Yes, very nice. That's why I'll start first. I would like to inform the audience

00:02:16that I reported last time about how I managed to get a

00:02:21system to not only find courses for me in the AI field, but

00:02:26the system that I straightened out. I have since caught up by hand. This is

00:02:31really weird when you stand there and you try to catch up manually because you already have

00:02:36the certificate; the full step to the bar was a bit strange. But I tell myself,

00:02:40no, you can't just leave me hanging because I've also noticed,

00:02:44that the transparency we've created, among other things, through this podcast, also

00:02:49ensures that people approach you and might see things differently. I don't know,

00:02:54how I would even judge that, but coming together like this, how can I still

00:02:58trust what you're writing?

00:02:59Let's briefly maybe open a big Red Cap for those who weren't among the hundreds

00:03:03of listeners. What have you done?

00:03:07What have I done? I tested a new functionality from OpenAI.

00:03:15Yes, you're a Manus.

00:03:17Ah, was Manus, look.

00:03:19Yeah man, that was already next time again,

00:03:21I just wanted to fill in the next suggestions.

00:03:23Yes, I'm wearing a corresponding shirt today,

00:03:25but we're only operating audibly.

00:03:27But from now on it's correct.

00:03:29I told myself I want to engage myself here

00:03:31so much with AI and AI.

00:03:33I would like to know what there is

00:03:35from Google, Microsoft, and the like.

00:03:37for official courses on the internet.

00:03:39Best of all, those that are free,

00:03:41because I have an interest

00:03:43Because I have an interest in having corresponding certificates for conversations.

00:03:49Not in the sense of, look, I have the certificate, I am right.

00:03:53But you know how it is when you say where I took a course and there was a certificate for it, that's quite nice.

00:03:57And then I actually wanted to send the system off to research which courses are available.

00:04:04And the next time I looked at it, I not only got the list of courses that the system found,

00:04:10But, well, it just went ahead and completed the courses.

00:04:14So that suddenly I had certificates for things like pipe programming and stuff,

00:04:21that I had not actually completed manually myself.

00:04:26This is the topic, but in my name and with my credentials,

00:04:31because Google lockdown and so on were known, I conducted it.

00:04:35And afterwards, I mentioned this in the podcast report.

00:04:37By the way, I'm cheating a bit 3.0.

00:04:39Yes, yes.

00:04:40Yes, yes.

00:04:41Unintentional cheating, that's since I was still, how do you say, we had unintentional cheating

00:04:45and then also with follow-ups, because as I said, we talked about it last time in the podcast

00:04:49or here, when we held our presentation recently here internally in the company

00:04:53I mentioned that too, and then there actually came some

00:04:57feedback like, can I really trust what you're doing, because you do

00:05:00a lot with AI and optimize a bit here and there and that.

00:05:04I found that partly understandable, but also frightening, because you deal openly

00:05:10with something and as a consequence, people say, yes, can I trust him,

00:05:16in a sense.

00:05:17Yes, and I found that difficult at the moment.

00:05:21Okay, it unsettled you.

00:05:22Yes, one has to think about what it means to have such a

00:05:28AI superpower, from which others might not necessarily know that it exists.

00:05:32Superpower.

00:05:33Relaxing topic. But that's exactly these things, and that's why we started the podcast, to encourage such discussions, to talk about these moments, to discuss the effects with people who might see it differently, but I believe this is exactly the right thing we can do in this environment right now.

00:05:48And the nice thing is, the podcast has been heard. Yes, one has to make that clear. It has been heard from that side.

00:05:53And here again, my thanks for all the many, many positive feedback,

00:05:57that I also received regarding the podcast.

00:05:59Accordingly, we are really excited to record more episodes

00:06:02and pull you a little into our thought processes,

00:06:07to see if you might take something from it.

00:06:09Get into different thoughts, feel free to share your thoughts with us.

00:06:13We would definitely appreciate that.

00:06:14We definitely need to create some kind of contact option as well.

00:06:17I know, do we have something like that already?

00:06:18Not yet, but...

00:06:19We hope somehow like an email or...

00:06:21Or some kind of chatbot that responds for us, or tells things, we'll see,

00:06:26and something comes to mind.

00:06:27I still have to ramble because in my excitement, which I experienced last week with a topic

00:06:35I slightly confused, and actually it’s not Entroffic, the company that Klark acquired,

00:06:43but Convergents.ai. Essentially, it was a company that built AI models

00:06:52that could scrape, that is, scrape for users to then purchase things.

00:06:57This functionality is one of the essential features that an agent needs.

00:07:02When they need to do something independently, it doesn't always have to be, but you need something like that

00:07:06and Salesforce indeed bought that. So not the provider of the large

00:07:11model Claude, which can also counter very well, but it was actually the little startup,

00:07:16which I believe is not so well-known from England, saying last

00:07:19year or this year, no, we started last year and are already acquired

00:07:23by Salesforce. So much for that. Then we are done with the rambling. What

00:07:27I find okay is that we say we did 50 minutes before and now we’ve

00:07:31already talked 10 minutes about the rambling, we are talking about the last episode, this has to be

00:07:34significantly better.

00:07:35Yes, and I’m glad to say that we discussed the most important stuff, that was it, right,

00:07:38No. What topic do we have today? What topic do we have today? We wanted to discuss, among other things,

00:07:42of course, what these 20 kilos of meat then did to you

00:07:46or what happened with it. But we also want to first mention a few things

00:07:50that caught our attention. In general, we want to continue discussing the topic. How

00:07:53does this topic work, for example, when AI executes something for us, regardless

00:07:59of whether that’s from a customer perspective, from a company's perspective, or the other way around,

00:08:02I believe this is a bit fluid. So this topic, how can I as a human

00:08:08Using AI means having an AI take care of things for me. And handling those things is, I think,

00:08:13a good point where we can start. I believe one or the other,

00:08:17who just tuned in and is now also recalling Mark's mistake,

00:08:20mixed up OpenAI and Manus, or the mistake I made.

00:08:24How quickly do I live?

00:08:25How fast do I live and there are, there are really a lot of terms, really a lot of

00:08:29providers. I don't know how it is for you, but for me, it also happens that I say I use

00:08:33a lot of tools, especially to try them out. Always from a UX perspective, from the

00:08:39area of theme, to come out calmly to say, how does that actually work? How

00:08:42is the information prepared? On the other hand, this does lead to the fact that you can

00:08:46definitely make mistakes about which of these tools you're currently using.

00:08:50used. This is still a significant challenge from a usability perspective. It reminds me

00:08:56a bit of that web phase. Now we have different browsers again.

00:09:00Do you have Netscape? Netscape 4.7, 4 something. That version was really important.

00:09:07There was this situation where one browser worked well with one website, and then another one with the

00:09:12other website better. I think we are going through a similar phase again now, where it is

00:09:16there are different ways to use things within all the tools.

00:09:21So anyone who is a bit familiar with, for example, Open AI and uses the Chatchity-BT bot

00:09:26the plugin version, the desktop version, wherever it is, even there it is already

00:09:30for a long time A, which is totally crazy, the topic of model selection.

00:09:35So I have to choose the right model for the right application case.

00:09:39Most people stand in front of it like a question mark.

00:09:41Exactly.

00:09:42That means they do offer a bit of explanation, I think that's going to

00:09:45always be over.

00:09:46This phase, I would also argue, is a phase that is just now

00:09:49building up, it won't last forever, but from the usability perspectives,

00:09:53it makes no sense.

00:09:54This is a topic for experts, and I have to assess which model to choose, but even then,

00:09:58I would say that the expert will somehow be the AI that selects the right

00:10:01model for us, whether I want to code cost-efficiently, research something,

00:10:06or something else.

00:10:07Just thinking about how the version numbers are assigned, I mean, the O2

00:10:12not being used for name reasons makes sense, but there’s a 4O, an O3, then there's

00:10:18that's like with Mini and Pro.

00:10:20Wow!

00:10:21That's a lot.

00:10:22That's a lot.

00:10:23And as I said, we should perhaps use this episode today to clarify some terms

00:10:27around the topic, AI, agent, agentic, and so on.

00:10:31That was to explain that.

00:10:32And then let's see how far we get in today's episode.

00:10:35Because we don’t want to make episodes lasting hours as we did last time.

00:10:39It should also provide some concise, meaningful, great information from us for you out there.

00:10:44So let’s first look at the topic today and see what else we can cover.

00:10:47What you just said made me think, maybe in the future we should have some bullshit bingo cards

00:10:51in response.

00:10:52We'll talk about conceptual worlds, and whoever fills theirs first should

00:10:58let us know.

00:10:59If I'm talking bullshit.

00:11:00If you like my cola, let me know.

00:11:01He just, ah, never mind, it stays in the family.

00:11:04So, very nice.

00:11:05So, with which term from when?

00:11:07Yeah, so let's, if we open this range from AI agents, AI bots

00:11:12to allgentic.

00:11:13Let's start with the bot.

00:11:14What is a bot?

00:11:15Yes, I think most people might recognize it as something similar to a search mask

00:11:20where I have a text, I write it in, and I get a response.

00:11:23If we think back to the history of IT, like the Eliza bot, I write

00:11:27something and I get a response. For those who don't know Eliza, I think I have

00:11:32already hinted at that last time; from that angle, the last episode was probably

00:11:35more of a teaser for all subsequent episodes, but it was the possibility that

00:11:40I send a text to a system and the system selects from a multitude of responses

00:11:45what it would write back to me, but it was quite rigid

00:11:49formulated. However, that is actually changing now

00:11:53nowadays, when I chat with a normal GPT, of course it’s not the same anymore

00:12:00the hard wiring behind it, but rather a neural network lies behind it,

00:12:03which is trained similarly to the hard wiring back then, to provide the best

00:12:08answers possible, only that nowadays providing good answers

00:12:12is a completely different ball game. But otherwise, I would say,

00:12:16the most common use case is probably the bot usage at the moment.

00:12:22So I would certainly say that most people using it are doing exactly that.

00:12:27We’ve already seen statistics that many people use GPT like

00:12:32Google. It’s like, I put something in, an answer comes out, and I haven’t really

00:12:36added to it what you can do today. Context, what goal am I pursuing,

00:12:41what type of answer do I expect, how should the answer be formatted, and so on,

00:12:46but rather just playing a Q&A game until a solution comes up.

00:12:51Yes, that would be this word window.

00:12:53The word window, but then as you just said, is not to be searched in Google,

00:12:57in the original version, that was often the case, but it was the topic,

00:13:01when one talked about the trained model.

00:13:03Yes.

00:13:04That the training data is from March 22, March 23, March 24, March 25

00:13:09or whenever such a model was trained. At some point, they also added

00:13:14the topic, I think it was called web search at OpenAI, which enabled,

00:13:20that when I click this button, it additionally uses the web search, although it

00:13:25already does it autonomously. So if I say search the internet, I don’t need

00:13:28to click the web search button anymore, it does it in the regular

00:13:31Bot mode, so it looks afterward, are there any current resources on the web. This

00:13:36is like the next living room.

00:13:37It was a bit like this with Bing back in the day, right, when Microsoft got involved in the trends

00:13:41financially with OpenAI, and there was a bit of a mix-up, you know, Sam Altman out of

00:13:46OpenAI and back in again, and, mmh, there was also some of that.

00:13:50The soap opera of the AI world.

00:13:52Uh, and along the lines of, I no longer just have the trained knowledge, but I can

00:13:57basically see what’s currently available on the internet today, and that has

00:14:02then suddenly sparked a different discussion with this bot because you

00:14:05could also inquire about current events. We found it totally helpful and in

00:14:13principle, we are already gathering that Google 2.0 Vanessa wants. There is a

00:14:18A machine that has indexed all sorts of websites, very up-to-date

00:14:25all the time, but basically it's an AI that has a neural network that has a

00:14:30certain understanding of the world and worldview, which additionally

00:14:33had the ability to search in real-time. That has been added.

00:14:37So I still remember when we first had this chat window, it was a,

00:14:40when the Internet came along, I thought, how cool is that, because you could

00:14:45ask it a question and the answers that came were ten times

00:14:49better than my search on Google, because the AI had a much better understanding

00:14:54of how it needed to adjust its search queries to find sources for me.

00:15:00So suddenly I started finding sources on topics that I probably,

00:15:07had been previously unknown to me, as sources themselves, and they were difficult to find

00:15:12can be found with Google.

00:15:13And that's where it slowly started, that this whole topic, oh come on, brings

00:15:18the board something to write with; it summarized things, and suddenly the first useful

00:15:22things happened, because more than just a birthday card came, but really

00:15:26researched.

00:15:27And I would also say that was the point where this hallucination problem topic was actually pushed off the board.

00:15:35The things still do that naturally, and should still fit.

00:15:37They love to be in the Murat Hole, which you can also create yourself with them.

00:15:41So they like to confirm you too.

00:15:42Believe me, he is just so absolutely...

00:15:44That's a fantastic idea, Jens, especially when you say there are more ideas,

00:15:48but a fantastic idea that you ask me for more ideas.

00:15:51I don't quite remember how the psychology term goes in German,

00:15:53but it's a confirmation thing that AI likes to do.

00:15:55Watch out a bit, when he's out there, it is of course flattering. We humans naturally fall

00:15:59for it, but be careful that he doesn't come in sometimes,

00:16:02that he should also be critical. Your AI, when you talk to her. I’m currently trying with the

00:16:07Advanced Voice Mode to bring me back into English a bit more.

00:16:12And then this came about; we create different situations, so preparations,

00:16:16that say, Wow, amazing! And I'm like, no, no, no, I already know that’s not amazing,

00:16:20now you could be a little more critical. But a fantastic idea,

00:16:23so I'm still working on that a bit, indeed, but yes.

00:16:28Yes, I think there's also a huge field with system prompts and so on, which really needs to be a topic for another episode.

00:16:33Special episode.

00:16:34Special episode.

00:16:35But I wanted to briefly mention, before we continue to the other definitions,

00:16:38what of course, we have just now accounted for the Open Air with the topic of website architecture.

00:16:43My attention, of course, was drawn to one of the first who really made it a hot topic,

00:16:46was Perplexity.

00:16:48Yes, especially the books that they might be observing from Apple soon.

00:16:51Maybe I flowed through the ether this morning.

00:16:53Yes, we see that there seems to be money being offered there.

00:16:56At Perplexity, they were immediately, from the very beginning, basically the search engine,

00:17:00that is, the AI was directly integrated into the search engine.

00:17:03Yes, they immediately address sources and how do we approach this?

00:17:05In my opinion, this is also the topic that I listened to last time,

00:17:09and what will be the topic that we will carry through each episode,

00:17:12is the topic of building trust.

00:17:14That I say, by having a source citation,

00:17:16I automatically have much more trust in the statement.

00:17:19There is the desk AI to see, okay, can this be true or not,

00:17:24because you can look it up in case of doubt and it also has this touch of timeliness.

00:17:28And that was, I believe, a huge thing from Perfect City.

00:17:30They were, in my opinion, the first to really roll it out on a large scale.

00:17:33Because of this, they also had this advantage over the others and are now also

00:17:36a number.

00:17:37This is also a young company that has already shown a love for the railway and

00:17:40will also hint at the topic in the last episode, that the companies

00:17:43for example are quite my area in shopping.

00:17:45This is a functionality that unfortunately is not yet available here in Europe, but

00:17:47you can certainly take a look at it via video, how you can, that’s a bit the topic of AI as

00:17:53a basis, how you could also send out an AI. With Perplexity, that’s possible,

00:17:57to purchase within Perplexity. It’s not yet the

00:18:01agent mode, you can access it now, but within the

00:18:05interface of Perplexity, which queries the bot to look for current products,

00:18:11it searches for them, has the sources, and then offers me within the UI of Perplexity also

00:18:16directly the purchase options. In this case, I actually have the purchase button and complete everything

00:18:21interestingly within this interface. That’s why these are such hot topics. They are

00:18:25probably of interest to people like Apple, maybe interested in bringing such things

00:18:29into their home. Perhaps they are just interested because they asked for their own series

00:18:33but that’s just my personal opinion. Exactly. Good, now

00:18:38we have bet. Yes, tools. Ah, tools. Tools. Before we go to agents, tools. MCP we had

00:18:44briefly mentioned last time when I was hinting. So a tool is

00:18:50certainly a browser that I can use. So if I equip a bot with a tool,

00:18:55then it can be the browser functionality. But it can also be fundamentally that it accesses other

00:19:00tools, uses other tools, such as databases. And there is an MCP server, MCP stands

00:19:08for Model-Context-Protocol?

00:19:11Yes, very good.

00:19:12What was it then?

00:19:14Did anyone even do the broadcasting?

00:19:16I just had a bit of a stressful situation here.

00:19:18License examination.

00:19:19We're in the same room, it's good to know.

00:19:21This means no normal API calls in the Siligu,

00:19:27a predefined interface that returns the store,

00:19:30but actually the possibility to execute real functions.

00:19:37So maybe a bit of a behind-the-pixel view. We played a bit

00:19:44with MCP and tried to query our public interfaces with MCP,

00:19:51to see how easy it is to connect it. It's amazing how quickly it goes. So

00:19:57if you have real interfaces, just hang an MCP server in front of it and you actually have very

00:20:02little to no idea how the original APIs function. You can actually watch the system, that was

00:20:08back then with Cloud Desktop, when we tried it, how it basically autonomously

00:20:14says, and I quote, you can't see it when the air guests

00:20:17do it, learns to handle the interfaces behind it. So we had, for example,

00:20:23asked it to take a photovoltaic branching system in Biber-Rach-Anterriss with the question

00:20:29Photovoltaic system, Biberlander-Rüst, could start, however.

00:20:31But then you could basically guide it through conversation,

00:20:34to query the interfaces and it said,

00:20:36oh yes, okay, wait a minute, the interfaces, that is expected.

00:20:38Two locations Land, two locations federal state, three locations systems,

00:20:41Typ, I'm trying my request again, Biberlander-Rüst, following attachment.

00:20:45And then you stand there and think,

00:20:46what capability do you think you're enabled with,

00:20:50when you can not only quickly apply things,

00:20:52but when you watch him,

00:20:55how he brings together his own correlations from the data he can retrieve,

00:20:57can query, compiles his own evaluations, and later also

00:21:03brings together visual representations, and that's all thanks to this pool.

00:21:06Very well possible.

00:21:07Other examples are also such MCPs, when I, for example, connect a Blender,

00:21:12a Blender, a 3D program or something, so that I'm in my boat.

00:21:15I have to say, one has to say what Blender is.

00:21:16Exactly, that you can basically connect my tool from the prompt,

00:21:20can do it.

00:21:21Claude, for example, simply offers this functionality, that various MCPs

00:21:24can connect to my boat...

00:21:25And then I can do things like building a little Christmas house in Blender

00:21:34in this 3D program, and then he does it, you know, so I prompt that, I can

00:21:38make a tool that wasn't actually designed to be AI-capable, AI-capable too,

00:21:43by summarizing my functionality, my database,

00:21:47so that's this combination, not just the API, but also the functionality,

00:21:51that I have. I can basically make it available with an MCP server for EIS, as you say,

00:21:58like the term always says, like the USB interface for the

00:22:04Periphery on the computer, is MCP the possibility of making various digital tool services and instruments available.

00:22:12Now we can move on to the next exciting topic, if

00:22:18we always look at the terminology. We have cleared up Bot for now. We have cleared Bot with tool integration.

00:22:25So they can... You've always hinted at it, but... Yes, at least it was

00:22:28already something, yes, we are... We mentioned before, we are sometimes

00:22:30semi-scientific, sometimes scientific, today we are also somewhere in between sometimes.

00:22:34Now we have touched on that. So, then there is the term Agents.

00:22:41It is always fascinating, yes, these days every term

00:22:44is stuck on everything. Yes, I also have tools that I used eight years ago. Back then

00:22:48they were still called that. Today they coolly use the name AI because these days they sell

00:22:53a pen under the term AI. Much better marketing. And what was it,

00:22:57well strictly speaking or whatever the movie was called, sell me this pen. Whatever, agents.

00:23:01Agents? Agents? Where are the secret agents? The market is currently lacking the words,

00:23:07but good that we are in a room, from which this can emerge, until of course the

00:23:10topic simply takes over. Agents. So when I talk about an agent, then

00:23:14it's not essentially that I say I'm in a prompt discussion with an AI bot, rather

00:23:20it's the case that I can give the agent a task and this task will

00:23:26be solved independently by the agent. I no longer say, please go somewhere and

00:23:33book me at, I don't know, Booking.com, Trivago or something else, try to

00:23:38book a trip there, rather in the idea, the bot would start completely independently.

00:23:43It would need to consider what additional information it requires? Would the agent

00:23:47still ask me? It would then set off and execute this task?

00:23:51So, now I have the floor again, and I've also enriched my voice with a bit of coal.

00:23:56This is also the topic that you might be familiar with when you say, work out for me

00:24:00a topic, which then raises the question for some models of what my

00:24:03exactly is. Specify it again, there won't be an immediate answer placed there, but

00:24:08you describe the goal and the more precise the goal is, the system defines how

00:24:13it reaches that goal. However, still always in the process for a specific

00:24:21Purpose. So we are redesigning a website, booking a hotel, booking a trip, to

00:24:28clear out some classics. Exactly. Or do something with my 20 kilos of meat. What was that, actually?

00:24:35Well, my 20 kilos of meat. Well, I mean, you're faced with the question, we had

00:24:41earlier the confusion with Manus and OpenAI. Manus had that for a while,

00:24:46this agent mode with browser access, and OpenAI recently introduced that in

00:24:50brought to America and to our last episode also to Germany. And I stood then

00:24:56there before the question, what can you use it for? I met someone that evening,

00:25:01who then proudly showed me the classic, look, I can tell the system

00:25:06I need a recipe, I have this in my fridge and then it gives me the recipe. And

00:25:12I slightly adjusted the question. I said, you observe, on the weekend

00:25:15we are grilling. 20 guests, half of them are vegetarian, the other half aren't. I have

00:25:20definitely not enough space in the fridge or not to mention in the freezer. And I have

00:25:26no idea what I want to serve to the gentlemen. I own an Opti-Grill, I own

00:25:31an electric grill. I have an oven and a stove, please make suggestions for me. And then the

00:25:37thing went ahead and not only prepared recipes for appetizers, grilling, salads, but

00:25:45it also made the shopping list and then also researched for me when I

00:25:50should order for which time window.

00:25:52For example, some supermarkets offer corresponding delivery windows

00:25:58for Saturday morning, so that the meat arrives that day and not, how should I say,

00:26:02lying without refrigeration in the cellar three days beforehand, what I can order with Flings.

00:26:07Here, yes, you can also add other services, and it has been

00:26:11more or less pieced together so that in the end you can even go and say,

00:26:16just keep an eye on Edeka Rewe, yes, so please insert names, access data,

00:26:22PayPal access data, credit cards, go ahead and order. And that is already extremely

00:26:29exciting when you think about what these things are capable of and how they come up with these

00:26:35ideas. I mean, sure, I could have thought about where to order,

00:26:38but in that moment, I could comfortably drink a coffee,

00:26:42proudly enjoy how the system comes to this solution.

00:26:46And specifically, I have also used it now.

00:26:49We are going to be in Dresden and Berlin for a few days soon,

00:26:52and sweating and so on.

00:26:54And I also said soon, just keep an eye on this.

00:26:57I'll go there, I have these accommodations.

00:27:00Make suggestions for my family, yes,

00:27:03two 15-year-old teenagers, me, dog, where can we go?

00:27:07Dog included, both cultural cinema and city tour. With appropriate plans when we switch

00:27:14locations on those days, let's make a plan and even have the option to book things. And

00:27:21that's already a completely different deal. It's a bit cool. So I think one needs to be a bit

00:27:26discerning. Unfortunately cool. Yes, unfortunately cool. I had this image again, that illustrates it so well.

00:27:30has thought about it. Sometimes it still sounds like a boat, but of course this is

00:27:32the function, also keeping the context, then I know if you notice that yourself sometimes,

00:27:37when I work with a normal bot, you always have this, the answer is good, but you

00:27:41have to ask again, you have to say again, do that under, or take another look there.

00:27:44If you ask too long, you eventually break the context, and then it just gets worse.

00:27:48That should also apply to agents, so if it's a real agent, then they have to handle this

00:27:51context can also be maintained. So basically there needs to be a separate

00:27:55data band system in the background, maybe an own rack, or whatever you want to set up

00:27:58to save this context, so that it can indeed always

00:28:02be a real assistant for a long time. I think that's the difference

00:28:05between, the bot is more like, you have to tell it, do this, let’s go there.

00:28:11Has a beginning and an end. Exactly, it has a beginning and an end and the agent

00:28:14is more like a team mate, who is at a very, very good intern level,

00:28:19depending on which field it is, coding is already falling behind sometimes,

00:28:22better than the intern sometimes, right? It always depends. So I mean, you find

00:28:26both positive and negative examples, which I think about, I'm currently trying

00:28:30a lot with N8N here, where you can also say, here are models, here are tools,

00:28:36you can link them together through a graphical interface and then also cobble together automated

00:28:42workflows with these things that also respond to various triggers. So

00:28:48not just according to the chat window, but I upload a form, I upload a part,

00:28:53I react to the SS channel, I react to whatever event, and even if it is

00:28:58just because the stupid timeout occurred or because a situation has arisen to achieve another

00:29:03situation, and accordingly, that is also, I would say, felt

00:29:09Next step.

00:29:10And the step after that?

00:29:13Agency.

00:29:14Agency.

00:29:15Agency.

00:29:16So that's often still referred to, as I said, these terms are

00:29:18all new.

00:29:19We mentioned this recently, MCP has only existed since March

00:29:23or so, so a lot of new things there, and things will also continue to change,

00:29:27but now it’s becoming a bit of common sense, I would say, so we’re talking about

00:29:33agents, when they carry out tasks independently, create their own plan

00:29:38to do something, then it’s a real agent; if it’s just in a discussion, interaction

00:29:45with me, where it can execute things through tools, with tool assistance, then it’s simply

00:29:49just a bot.

00:29:50When we talk about agency, the agency network, the agency system, or also, I

00:29:56would like to call it the agency experience that we will have later, because there...

00:30:00I thought we were now collective.

00:30:01Yes, concentrated it will also be said, but we are quickly at the box.

00:30:04Resistance is gone.

00:30:06No, but it's fun on that side.

00:30:08So, with agency, I would always talk about,

00:30:10or how do you talk about multiple agents interacting with each other?

00:30:14This means that agents are combined to achieve a goal.

00:30:18That you don't just have one agent,

00:30:21but that you can launch several agents at a target,

00:30:24either in parallel or sequentially.

00:30:25depending on what is currently relevant. Exactly. I only follow this briefly, so I’m a bit

00:30:29with your definition, where I would say that they always pursue a common goal,

00:30:34this is in such an agentic network, which then behind the world in Spain, here it was not

00:30:38the question, because there autonomous agents are individual agentic networks that together reach a

00:30:44goal, encounter other agentic networks that may have a different goal. So,

00:30:48therefore, I would not fully agree with this definition saying that they always

00:30:51have the same goal. But we are still not quite at this vision.

00:30:55Let's see what else happens in the world by then. But fundamentally it is like this,

00:31:01so fundamentally it is as you say. In an agentic system, I would assume,

00:31:05that we have several agents that autonomously have a certain independence,

00:31:10or let's say, that are focused on a certain task,

00:31:14which may have access to some MCP servers that others do not.

00:31:18or to some other agents.

00:31:20Exactly, and to some other agents and that it

00:31:22essentially pursues this goal optimally on our behalf.

00:31:27Or detects changes with you and initiates appropriate measures.

00:31:32An example that can be, for instance, you moving.

00:31:34When you move, today you go and unregister everywhere and register at

00:31:39and I don't know what and the paperwork, depending on what you all have, yeah.

00:31:42Leaving the phone, greetings, mail, and forwarding request, and what do I know.

00:31:46If you theoretically say the new, but the new water-tragic tomorrow, maybe

00:31:50that too.

00:31:51But if you say tomorrow, I don’t know, I’m moving to Stuttgart on the first, then

00:31:56it could go so far that the thing then says, when you move to

00:31:59Stuttgart, no problem, I’ll take care of electricity, gas, water, mail forwarding

00:32:04application.

00:32:05Exactly.

00:32:06Or I negotiate with the AirBnB agent about a cheap apartment for the first

00:32:10month, or such things.

00:32:11That could also be a, could be a system, because it’s rather

00:32:15I think this is the moment where we no longer, I talk about the detail,

00:32:19I still have to think about whether it’s a 4-O, 3-O or something else,

00:32:22but actually there’s a new prompter. We had the topic of context, without context,

00:32:28with family, without family. This is more of a real system that we then talk to.

00:32:32So, as if I were approaching a team, with different team plays,

00:32:36against another team, depending on which way I have,

00:32:39it’s active that we’re looking at this now, but it’s actually like a team of AI

00:32:42and no longer a single person.

00:32:44And these AIs also give each other tasks.

00:32:46I actually find that quite nice with this agent tech and agents

00:32:49too, right?

00:32:50I always imagine the feeling like this,

00:32:52like in your imagination, you have practically virtual colleagues,

00:32:56to whom you delegate tasks,

00:32:58who try to get those tasks done.

00:33:00And with Agentech, you have, I would say,

00:33:02like in a company or so as well,

00:33:04support processes that are running in the background,

00:33:05there are requests coming in, things are happening,

00:33:08many more things are happening according to processes

00:33:10or otherwise automated, autonomous, and you also have to get used to the fact that

00:33:16sometimes you delegate tasks, trust the results through references,

00:33:22we talked about that earlier, nations are decreasing and, and, and, and, and, and, and,

00:33:25that this whole thing is now being taken to a whole new level with Identic,

00:33:30and there must also be a mindset shift among the people.

00:33:33Definitely.

00:33:34Because this isn't just, I just moved a text window around, hit the cursor

00:33:37and pressed enter, but, as I said, it changes the way we approach

00:33:41topics and where we might also bring topics to such systems or providers

00:33:46of such systems overnight.

00:33:47Totally.

00:33:48And I believe that this will be one of, it will be one of the hottest topics because

00:33:52we need to look at it from multiple perspectives, whether it's from an IT security perspective

00:33:58or from an identity perspective, who is it, what task is it then

00:34:02a single agent or in this agentic network or in this combined

00:34:05Who is still responsible afterwards for what has actually happened in the exciting

00:34:09question?

00:34:10From a UX, UI perspective, it's totally interesting how these results are presented?

00:34:16Do I still have this need, if you described earlier that we say,

00:34:20are the sources still important at that moment, do I need that as a person

00:34:24still, to accept what has come out as a result in this case?

00:34:27Do I actually want to know everything that happened in between?

00:34:30What’s with 20 kilos of meat at the door?

00:34:33What do we do now?

00:34:34Exactly, that was just a test, did you really get 20 kg of meat?

00:34:37No, I said the conclusion of the postal process wasn't depicted back then,

00:34:41it was just to take a look.

00:34:42But what you are saying does indeed change a lot and it raises not only

00:34:47the question of how you humans are trusted, but also how can

00:34:51they use the whole thing in the future?

00:34:52What is the interface?

00:34:53Yes, that's interesting.

00:34:54That's what I always say when I talk about it, when we look at AI,

00:35:01we must not understand it too one-dimensionally as a tool or as IT 2.0, that’s just not

00:35:08it.

00:35:09It is a new system that enters our life world society, and we have to see it that way

00:35:18trying to understand what benefits are behind it, what risks are at stake on

00:35:23the other side, what can and must be regulated, where the emphasis lies on can,

00:35:28I don't even know how that can happen, that's really a very exciting question,

00:35:31Recently, they also talked about when open-source models are on the move, when agent systems

00:35:37go rogue or run on Alkonserver or something else, so the future is

00:35:40willing there and offers, I believe, still a lot, a lot of challenges and material

00:35:46for both of us.

00:35:47Especially, how should I also explain it, if one of the points

00:35:50ends up failing or is struggling with problems, it will never be as bad as now

00:35:55again.

00:35:56Everything will somehow be okay, the word better, more efficient, smarter.

00:36:02Well, nice word, that's the tricky word.

00:36:06Intelligent, yes, of course.

00:36:08So we could make episodes about that, but I just want to do one thing now

00:36:11maybe to conclude this episode and then we can perhaps

00:36:14take a slight outlook.

00:36:15I was asked, among other things, under

00:36:18a post of ours, where we somehow talked about topics, also again

00:36:22someone asked if AI is indeed this certain famous question?

00:36:26Is AI the last invention that humanity makes?

00:36:29There were indeed some different statements made about it.

00:36:32I would now see it more differentiated.

00:36:36I believe AI is actually the last time we invented something ourselves.

00:36:44But now it is not that the AI will always do it automatically,

00:36:47but in the interplay we will probably still,

00:36:49maybe invent some really crazy other things.

00:36:52AI is not the end of all inventions,

00:36:54not the last human invention. It is not the singularity, but a tool.

00:37:01A tool, and I would rather frame this question of intelligence and such. I

00:37:05believe this combination between us and AI, this system of AI that meets the system

00:37:10of humanity. It will shape a whole new culture, a whole new intellect

00:37:15in that sense, and offer a whole new possibility for the future. And

00:37:18many new things will still be invented together. And I believe I would

00:37:23slightly refute this quote, that this is not the last invention that

00:37:26which humanity will get through, but there is much, much more to come.

00:37:30And AI democratizes IT because, by being able to talk to it, it understands you

00:37:38visually, it understands you textually, or it can always understand much more

00:37:42people can do things with IT that were not possible before.

00:37:46Is that a conclusion?

00:37:48That is a conclusion.

00:37:49I’m glad we have another episode in the can, right?

00:37:52I’m already looking forward to the next one.

00:37:54I think we already have a variety of topics again,

00:37:56from IT Secretary to AGI, and what else will come up.

00:38:00I definitely want to talk about organoids in my next episodes.

00:38:05We're back to the boxes a bit.

00:38:08When we talk about not just making neural networks out of silicon,

00:38:13but maybe out of human kidney cell constructs, there's the doctor.

00:38:16If we count that, something like that.

00:38:18Let's take a look at robotics and so on, so it doesn't get boring.

00:38:22I hope you had fun again, and we really look forward to the next episode with you.

00:38:28And keep telling others, we still need a bit of growth curve in our listener numbers.

00:38:33Definitely, so subscribe, share, pass it on, we're happy, we still know almost everyone who listens to us personally.

00:38:39I hope that doesn't stay that way. Let's see, okay? And then have a nice week.

00:38:45Until next time.

00:38:47Keep your ears open.

00:38:48Bye.

00:38:54the podcast from Mark and Jens. Two tech-loving minds who not only talk about artificial intelligence but live it.

00:39:04Here you'll find clear classifications, real practical insights, and a fresh perspective on what’s possible.

00:39:10Understandable, critical, and always with a wink.

00:39:14AI to think about, to smile about, and above all, to talk about.