Clawdbot

Transcript

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

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

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

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

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

00:00:29A week has gone by, and I welcome you together with Jens to a new episode of Think Different,

00:00:39Think AI.

00:00:40Today it's just the two of us, but we still talk about an imaginary third person.

00:00:47in the room, namely a real personal assistant, as one might recognize from

00:00:54Jarvis from the Marvel universe. But in truth, when you look back at it,

00:00:59maybe it's such a thick-headed series or whatever it still is nowadays,

00:01:05it's also funny when people see themselves, they always think,

00:01:09God, who knows how good or great they look when others see them,

00:01:13sometimes you think, well, I'm not quite so sure. As I said, I'm not alone yet.

00:01:18Jens is also here. Jens, the topic of personal assistant. I would say we have

00:01:24prepared a little something, let's start at the very basic level.

00:01:27The lowest level of the personal assistant, hello everyone. I'll say,

00:01:33a personal assistant is something that we've wanted for a long, long,

00:01:37long time. As soon as we started having conflicts about this topic and thought

00:01:42okay, these machines should actually be able to do more than just some calculations,

00:01:48write text, and do something else, but they should really just handle tasks

00:01:52for us. And I think the most basic level, I mean, whoever has tried this personal assistant touch

00:01:57in recent years was the wonderful Amazon with their

00:02:03Amazon Alexa, which was around the house as a bit of a personal assistant,

00:02:07or also Apple, with Apple Siri, the personal intelligence that you carry around on your phone,

00:02:14Both have now, let's say, landed on the most basic level, up to now.

00:02:22I like how you define the most basic level, I wanted to let you finish speaking,

00:02:26how you break down the most basic level. I thought of terminal windows as the most basic level,

00:02:31The text windows of our computers is also a bit of what I would say has been last time

00:02:39tainted by social media, namely I actually wanted to get to the knot point.

00:02:43Oh, are you already that far? I thought we wanted to start here first.

00:02:46Yeah, yeah, that’s okay, then. That’s a completely different level. I mean,

00:02:52where I just spoke about is actually such attempts

00:02:55to help us with devices in our households that partially try to assist us in daily life with

00:03:01voice and voice input and output.

00:03:07So, just classic things like ordering things or I can ask the device

00:03:12what the weather is like outside, whether it does or not.

00:03:14Just say, just think for yourself.

00:03:17You can figure it out yourself by looking outside.

00:03:19Exactly, look out the window, exactly.

00:03:20But that was always just on this very simple level to take off a few tasks.

00:03:26So, these devices have always failed when it came to that and then comes,

00:03:30I think that fits. And you just said now to play slowly. When it comes to something like this

00:03:34as simple as booking me a flight. Or anything that might still be quite simple

00:03:39probably, anything that somehow combined two topics wouldn't work with

00:03:45these assistants, like the earlier ones, at all. So something like, oh, if

00:03:50my dentist appointment tomorrow gets canceled, then book me a table somewhere in a

00:03:55restaurant where I can go or something like that, such things, like dual connections.

00:03:58If the dentist appointment gets canceled, I'll go for lunch.

00:04:00Exactly, I can go for lunch.

00:04:01I think dessert is delicious, right?

00:04:03So, that just doesn't work, right?

00:04:05And there, Mark, of course with that,

00:04:09that's already a gigantic leap from the one - and that just,

00:04:12now onto your - and that just, right?

00:04:14There is of course something with Claude Court, a,

00:04:17actually as a programmer, counterpart of Claude,

00:04:22conceptual AI tool,

00:04:26One can rightfully say that this is a real thing that exists and that

00:04:32a lot of things can basically be started from a Terminal application just like that,

00:04:36can be started.

00:04:37Perhaps to clarify, there are a few things you can do with Cloud Code.

00:04:41We have to pay a bit of attention to the naming because later we also have

00:04:44something that sounds similar, yeah, one really needs to speak slowly and clearly

00:04:49Let's talk about what we can do at Phänomenal.

00:04:51Yes, little dreams.

00:04:52I'm already excited.

00:04:53I mean, if someone else is listening, I'd definitely think about booking a moment and stuff like that,

00:04:57didn't those two old men already tell us, with Manus and Gentik, about Atlas Browser?

00:05:04Yes, that's true.

00:05:05Claude Kot was also out of my field of vision for a long time because I was distracted by the word Kot.

00:05:12Not because I have any fear of terminal windows, but it was a blocked connection and I thought,

00:05:17okay, fine, I program very little in my everyday life.

00:05:21Let's leave it alone. At some point, I noticed that various

00:05:26videos were circulating on the internet, showing that people could start cloud code, i.e., at terminal level.

00:05:32You can tell it to do things in the directory it started in,

00:05:37that it can plan projects there, write HTML pages, program games,

00:05:41write software, whatever. And that you can continue chatting with the result.

00:05:46But if you're already at such a terminal level, the thing can also do completely different

00:05:51things.

00:05:52Yes, for it, it doesn't matter if it works with a Word file or with

00:05:58an HTML file, that is, with a file that opens in Microsoft Word or with

00:06:03a file that can be opened in a browser.

00:06:05And there were very nice cases to see, and I tried it out then, I must

00:06:11admit to my shame, yes at this point greetings to those who are now quietly

00:06:15chuckling to themselves that I didn't manage it. I wanted to install it.

00:06:19Installation worked. You could also tell it to do something or you would have told it

00:06:22can say, you check the browser. It’s also great when you do that.

00:06:26But that didn't really work out. Was I somehow too stupid to figure that out? I should have

00:06:31just installed a plugin in Chrome for it to work.

00:06:33But well, I just didn't do it. Long story short, you were in

00:06:38a position to use it and get it done on your computer.

00:06:42And well, as it usually goes, it then somehow went, I’d say, pretty steeply uphill regarding the topic

00:06:50of visibility on social media, and apparently Stropic noticed that too and then

00:06:57offered things to people like me, who then, oh, I start in the terminal and get the browser to fly,

00:07:02in their Mac version of their cloud application, the topic of Co-Work.

00:07:10And Co-Work is simply put, not a terminal interface, but an interface from within their application.

00:07:18There you see, okay, I can define tasks, I can share a folder that they can work in.

00:07:25And then the thing starts and also takes on activities there.

00:07:30So what did I tell him, for example? I said to him, you check it out, I got myself a 4TB external drive.

00:07:35From old days, which is, let's call it, yes, genius loves chaos and that was the

00:07:41filing structure that I seemingly had on this disk many, many years ago

00:07:45I chose it and I connected the record and I said to it, for example

00:07:49can say, you are allowed to access the hard drive, organize that for me.

00:07:53And then the thing started, it laid out a plan, also made a

00:07:57structure, changed file names, marked files with

00:08:02similar content, and basically organized my hard drive, and I found

00:08:08that extremely impressive. Yes, you give it a task, walk away, and eventually you come

00:08:13back and then it’s working on your computer and then you try a few things

00:08:19and at some point I made the mistake of not telling him to reorganize,

00:08:24but I went over and said, here is my computer, here

00:08:28is my download folder, here is my desktop. Optimizing.

00:08:31And well, what did the little creature do? Well, not little creature, that's just my diminutive form

00:08:37for such things. He didn't just start but also created structures and files

00:08:41and added files, but I only saw that in the log files,

00:08:45he started and updated files for me. Saying, hey, this can’t be wrong,

00:08:50I can finish this. I'll write this for you. And here, I’ve updated an old

00:08:54one for you. Saying, a small error cropped up. I felt a

00:08:59bit reminded of the example with the paper clips, suddenly making sure,

00:09:02that I’m human in the loop, that I can now check, okay, I'll go over

00:09:07my log case, because such a file system is not versioned, it’s,

00:09:11what's done is done. And that's funny and you’ve already unmuted the microphone,

00:09:16there's one more thing I want to share, and then I’d like to pass the ball back to you,

00:09:19namely the topic of Claude, which also has something called skills. So you can

00:09:26train a skill and thus teach things. And with Claude, it’s such that you can

00:09:31not only write a skill that is like a markdown file where it's written,

00:09:35do this, then that, then this, pay special attention. You can also include things,

00:09:41that you might not want to leave to the AI because they are simply static and programmable,

00:09:46you can place code in such a skill, for example. And then there are scenarios that

00:09:51he really processes as program code, without interpretation. Other things, he then handles

00:09:57with his LLM, which is, I interpret it with my non-deterministic system,

00:10:02what I do with it. And also this skills are available to you not only in the classic cloud, but

00:10:08then, for example, also available in this co-work or in Cloud Code, and he can

00:10:14either pull them independently or you can instruct him to pull this for certain things

00:10:18from the Guild. The thing is so powerful. I found that incredible until 24 hours ago,

00:10:24but we'll get to the other part later. Okay, but that's before we go any further.

00:10:28And the last 24 hours, or maybe for some 48 hours, because over the weekend,

00:10:33the internet has been a bit wild. Let’s dive back in. What is currently

00:10:38being talked about, do you remember that phase when we also looked at how

00:10:45the usability of the models and the chat models had suffered a bit,

00:10:50because a tremendous number of new models came out? One had to consciously decide as

00:10:55a user suddenly, do I take the ChatGPT4O version or just the 4 version? That’s now

00:11:03God's sake, exactly, that's a bit behind us now, the providers

00:11:06have partly gone back. While the selection is still there, one could still select the models,

00:11:11but basically many providers have now decided, no, we let

00:11:17the model decide itself because it makes a reasoning. That was on the fingers,

00:11:20so auto mode, I don’t have to select it myself anymore.

00:11:23Now it’s something new again because the terms you just mentioned are many

00:11:26topics that are coming together now, and I say, even my head is smoking from it.

00:11:30And if you say now, we have the Claude model from, well, Claude

00:11:36application, let’s put it that way, from Anthropic with different models, whether that's the 4.4.4.5 so

00:11:44net or whatever they are all called. Net opus. Exactly, opus you didn’t see, the

00:11:49ones that sometimes have different strengths. That’s one thing, the classic

00:11:54LMMs, one has to say now, the cool things in the background, these giant models,

00:11:59they understand the world and can talk to us quite humorously. And then there is also for

00:12:05me explained again, there is Cloud Code. What is the difference between the

00:12:11normal Cloud application and the Cloud Code application? Yes, with the normal Cloud application you say,

00:12:18I want to get a text written, make me something, while the

00:12:22Code application can, for example, create project structures, can create folder structures,

00:12:28can work with entire folder structures that are also located with you. You are not uploading

00:12:33, for example, a large zip file and eventually get a zip as a response, but

00:12:38all files that are created are located on your box. And you can then basically

00:12:44also take care of versioning and such stuff yourself. And I think that is a big

00:12:50difference to that and first organizes his work package. He makes himself nests, how working.

00:12:57Exactly, so basically the normal Cloud application is more like this classic

00:13:01I am a chat thread and I actually work within this thread, where many

00:13:06things can happen.

00:13:07Applications can also be written there.

00:13:10So a particular incident was relatively early, where basically within the

00:13:13chat thread, real applications were already programmed back then by Cloud.

00:13:18And code can actually create proper structures on the computer and

00:13:22that can then be adopted.

00:13:24Then we had the topic of MCP servers, which essentially offered a way of access.

00:13:32So as a provider of something, I could make my application MCP-ready and thereby grant a model access.

00:13:41This already brought up the topics, that I choose Blender as a 3D program for example,

00:13:46that I can access both from the Cloud itself, from Cloud Code, but also gladly then essentially on these MCP servers

00:13:52And I can simply say to my model, build me a 3D castle and then it constructs that castle in the Blender tool, so from real 3D modeling.

00:14:03That was the topic of the MCP-Cyber, because the MCP-Cyber then provided the various functions in the AI program from Blender.

00:14:13Now skills. What are skills then?

00:14:18I like how you’ve once again structured my speaking overflow very well.

00:14:24I thank you for the help; someone has to be here to talk about structure and organization.

00:14:30So skills are, actually, skills are not magic, but just the documentation of,

00:14:39how to work on a sort of task description, and it's simply called skills.

00:14:46You could also just say it's a prompt, because in the majority of cases, that’s all it contains.

00:14:52With the difference that you can now use skills everywhere,

00:14:57that with Claude you still have this programming option, that you can say, pay attention.

00:15:02There are topics that you really handle with programming code.

00:15:06And with that, it's truly every time one plus one equals two,

00:15:10whereas an AI, if you just work with a prompt,

00:15:14can sometimes come to a different answer or a different formulation.

00:15:18And all of that together you can bundle in a skill package, save it

00:15:22and the system takes this skill package either upon instruction or based on its own decisions

00:15:28and evaluates it.

00:15:31And if you mentioned MCP earlier, that can also be mentioned,

00:15:34you can also specify which MCP service should be consumed by the skill, for example.

00:15:39Yes, so you can say, you have Mauber, and now you have something to do with the topic, you can use an MCT service ABC.

00:15:46A side note from me, you’ve all heard that I automate a lot with N8N, you can also offer MCT services with N8N now and let them be utilized by Claude.

00:15:57Okay, that would mean, that would mean, now to stay with our example from this 3D topic, I could do something like

00:16:04write a skill that I store or that I must load with the one focused on the topic of 3D modeling where I say,

00:16:12how he should behave as a 3D modeler, which lights to use, and.

00:16:15Which software, which MCP servers, so which 3D programs can access it without having to keep saying it anew, so I basically just need to provide him with this skill, and all this important information is then delivered.

00:16:27Okay, understood. I think it was important to clarify that, so thank you.

00:16:31So thanks first of all for the explanation, it really helped me, sometimes it can be a bit wild,

00:16:36but I think it's also important for all of us, there was always a bit of

00:16:40pausing at the speed that these topics have and trying to understand a little deeper,

00:16:45so what actually happens there. So, now we have that, you basically have, with Claude-Kort,

00:16:52I say this with this term, I’ll be exciting tonight, so you have an application,

00:16:59that can access the network through MCP server skills, which it additionally has,

00:17:05as well as access your local computer. It can handle all kinds of topics,

00:17:10that you enable. You just said, I can share folders with it. I suspect,

00:17:14you also shared your Google account or something else with him so he can

00:17:19send emails or something else, knowing you. I want to comment on some things

00:17:23not. But for example, I've also shared those mentioned MCP stories

00:17:26with Chrome. Is there a nice plug-in that he can use to control Chrome?

00:17:30Okay, so the elephant can get there so I can stay consistent here.

00:17:35That works with Cloud Code. But I only got that to work with Cloud Codework,

00:17:41because it gave me an error message I could do something with.

00:17:44Cloud Code hadn't done that. But I'm handing the structure back now.

00:17:47Okay, then it was just another conversation.

00:17:50Cloud Codework compared to Cloud Code.

00:17:55So now it should be said that probably five people are going to say here,

00:18:01well, with Cloud Code you can do this and that and that much better than with Co-Work.

00:18:06It is said that Cloud Code is for developers, Code Co-Work for knowledge workers,

00:18:11it basically offers me the possibility in the cloud app on an Apple Macintosh to say,

00:18:18in this directory you can do the following and I can also use the skills there,

00:18:24I can also have the parser controlled so that it does something there and then loads this directory for me.

00:18:32Or I can just use the normal functionalities of the models to describe task packages and have them run automatically.

00:18:43So I don't have a regular job, nothing where I can say, like, do that

00:18:50every morning at five, the code doesn't do that in Co-Work, but still it is

00:18:56very powerful because you can also go ahead and say, you know, I

00:19:01have a whole bunch of texts, could you do some proofreading, a little

00:19:05editorial skill, or research, the content is missing, or like in my

00:19:10example, if you express yourself a bit clumsily, it could also be that he gets a bit overzealous and then, yes, does things.

00:19:18Okay, good, understood. So, now we've sorted that a little and can take a look at the...

00:19:27We are now in the year 2026 on one of the last weekends, no matter how safe, probably the last weekend that January looks,

00:19:37Because another player appears from a developer from Vinesis, I believe, who has put this

00:19:44thing together for us.

00:19:46It's called Claude Bott, spelled C-L-A-B-D-B-O-T, so for everyone who wants to better understand behind me

00:19:58knowing how to pronounce it helps us, and then sends us a new message

00:20:02with the better pronunciation.

00:20:03Mark and I have already discussed in our preliminary talk how these different

00:20:07things actually sound phonetically different. But we really don't know for sure. It is

00:20:11also irrelevant. So there is a new player that is redefining the lowest level of personal assistance

00:20:20now because this is actually an application that, unlike a

00:20:28Chat-GPT is not running in a normal cloud in a chat fence, but can actually run locally on

00:20:34a computer and can fully access this computer

00:20:38very autonomously around these topics that we have just hinted at

00:20:43I give the cloud code, for example, skills so it can do something when I

00:20:50understood it correctly and I like that you have already played around with it and I

00:20:53have only researched, then it is the case that the cloud bot can actually

00:20:58he teaches himself all these things very very well so one really needs to

00:21:01then you only have one task left to give and then he starts and takes off.

00:21:05So, maybe we should also briefly address how

00:21:12married, because you mentioned web and it isn't a chat, how the little creature generally

00:21:16behaves or what is so special about the little creature, and then we will also come to

00:21:22Let's talk about the topic of skills.

00:21:25So Space Lobster is the idea, that is, a space lobster, yes, that's

00:21:33the iconic little creature we’re just discussing.

00:21:39You can install the piece soft, let's put it that way, on your computer.

00:21:46You also have now, I would say, noticed through this whole social media thing,

00:21:51that apparently Apple had a good weekend, because people seemingly

00:21:55thinks they urgently need a Mac for that. But you can also install it on

00:21:59a Raspberry Pi. Yes, so you don't necessarily have to take a Mac,

00:22:04even though I personally leaned more towards the Mac simply because

00:22:10my Raspberry wasn't that accessible. It somehow took too long for me. Yes, my Mac,

00:22:13was definitely easy to access. Yes, you get it. What can I say now?

00:22:18Yes, because I was just briefly about the hardware. I think,

00:22:20is important to understand. So I believe this software is designed to run on the

00:22:26Raspberry Pi or on any virtual private server you can get.

00:22:31So you don’t need to have a powerful machine. If you want to run it locally

00:22:36with a downloaded LMM, which

00:22:43has certain capabilities, then it makes sense to work with a naturally

00:22:48reasonable machine and a Mac Mini is ideal as

00:22:51a compact device that can also run Tandy Force 7.

00:22:54That’s why this thing is somewhat hyped.

00:22:56The actual software because it usually first accesses, for example,

00:23:00a cloud API again to be able to perform tasks.

00:23:03That runs on the minimal person, that’s just how it is.

00:23:07nice.

00:23:08There are also quite nice, small applications already with kind of respiratories and small

00:23:11monitor included, where then this comes into play in life and does all sorts of small

00:23:15things in the background.

00:23:18It's important to assess this, so everyone who wants to research what we are currently

00:23:23talking about and wants to try it out themselves, think about whether you order the

00:23:27Mac Mini directly.

00:23:28I checked the pulse four times over the weekend, researched several times about his

00:23:32roof.

00:23:33Then I'll first do the episode with Mark and see how we can categorize this

00:23:37and then I'll also think about whether I still get the Mac Mini or not.

00:23:41Yes, anyway you in Selysidas and then you're able to communicate with the thing

00:23:48basically via the messenger service you already have, whether it's Signal, whether it's Telegram,

00:23:55whether it's SportsApp, whether it's iMessage, you can then assign the computer some juice

00:24:02and interact with it, not because you open a browser window, but because

00:24:05it's basically operating within the system, in the chat communication that you already know, and there

00:24:12it waits at your command. As you said, you can integrate cloud AI models,

00:24:20that serve as the connections for its brain; you can, if you have sufficiently powerful

00:24:25hardware, also install local models and run them on the device. And yes,

00:24:35ultimately, you actually have the best of all worlds. You have a bit of this theme,

00:24:40which we also have from Cloud Code and Codework, but without the limitation of a

00:24:45party system into one folder, because you can give the thing access to the entire machine with everything,

00:24:52that is somehow on it, and because you just mentioned skills,

00:24:58there are skills that you can load. But you can also tell it to try to

00:25:02acquire the corresponding skills by itself and depending on how much it learns over time

00:25:08or installs accordingly, you can, because I just said, you can chat with it using the messenger apps,

00:25:14you can also send it, for example, voice messages and it

00:25:17transcribes them and sends you back voice messages. Yes,

00:25:21yes, you can converse with it, for example, and it learns, basically, while

00:25:29you write with it what it should do, from internet research to text creation and document generation.

00:25:39You can also give it tasks to be active on the internet, as perhaps we have also

00:25:44learned from Manus or now the plug-in integration from Claude, that you tell it,

00:25:51'Hey, keep an eye out and check here for travel planning, appointment planning on the internet,' and

00:25:57in the process, it essentially teaches itself things when one asks it to learn things.

00:26:03You need to hook back in, I'm going to have another flow of conversation when you say

00:26:07to teach things, then for example, you mentioned earlier the

00:26:09example that someone wanted to make a reservation for a table at all,

00:26:12and if it’s not even possible to get a table, and you

00:26:17have access to something like Eleven Labs, then you can also empower them to do that,

00:26:23you make an outbound phone call and say, just call them and make the appointment for me,

00:26:28so that on my, I don’t know, wedding anniversary or whatever it is, I have the opportunity

00:26:35to go have a meal in that place, and the funny thing is, you can really

00:26:44give it targets and it works through them and it especially remembers everything you’ve

00:26:50done with it. It creates Markdown files about all the interactions you had in the

00:26:57past. This means you have a kind of endless memory, so that it

00:27:04Thing that learns more and more not only about you over time but can also draw on it,

00:27:10how you react, what the right answer is when you need to write an email now,

00:27:15what you have access to, it's basically building a kind of persistent memory.

00:27:20Which is, of course, an essential characteristic for a personal assistant,

00:27:24if we take a look at today's topic again. And I would even

00:27:29to go that far, because you're saying that with 11 apps, 11 apps for the listeners out there,

00:27:36is basically one of the best voice AI model providers. You can wonderfully clone your own voice here.

00:27:43and stuff like that. And their sight is not good. One needs to have about 30 seconds of voice snippets

00:27:47upload themselves and can then really reproduce their own voice and all

00:27:52also all other possible voices. This is therefore also useful for many code center labs and so

00:27:57is a totally exciting application because you can naturally use it for voice agents.

00:28:01Happy listeners have already heard that because our Halloween episode and our Christmas episode featured stories told by voices from 11 Labs, so for those who don't know it yet, feel free to listen to the old episodes, even if Christmas is over and the holiday spirit might not be there.

00:28:19It comes back, it comes back every time.

00:28:21Yes, everything comes back, everything comes back.

00:28:23But what I wanted to say is, as I assess Claude Bott and also in the power

00:28:29this is, if it had access to your, let's say, a budget, then it would

00:28:34be able to carry it out independently.

00:28:36So, this thing seems to be really smart in its approach to problem finding

00:28:42and indeed through this strength, that I'm not just bound to some random

00:28:47chat window, but I really have access not only to the computer,

00:28:53so on the browser, I have access to the entire computer and can essentially use the whole computer with it,

00:28:58to do things, whether it's virtually lying somewhere or essentially then locally, and that's what's exciting about it,

00:29:05because then I can also build my little personal assistants a bit away from the big players out there,

00:29:10which can really operate quite autonomously.

00:29:16Yes, and find its own solutions, so that it is right now dealing with the skills or something, well, it would write them by itself.

00:29:23Well, so Claude would start writing skills and other topics by itself and continuously seek ways to work even if, for example, no MCP server were available, it would try to keep running until it found some solution for that.

00:29:36maybe to explain the power, yes, it is for me the first time the

00:29:42possibility, I mean, I have taken it, yes, why I have it, if I

00:29:48knows that no single Windows or portable device is in my private personal

00:29:53environment, and even though the software is still very fresh, it allowed me to access mail calendars

00:30:01task lists that are actually not really known for that.

00:30:08that they are easily consumable by some AI services. This means, one

00:30:15will probably, I can certainly talk about that in one of the next episodes

00:30:18whether that worked out well or not, also be able to do things like

00:30:22please check my emails, please take a look at a calendar. What does

00:30:25today actually have on the agenda? What do I need to think about?

00:30:28Make sure you define tasks that are basically on your task list and if he feels

00:30:33addressed by that, now he could even work them off. Even so

00:30:37far that can go. Yes, you enter something in your iPhone’s reminder app and say

00:30:42to him, maybe it happens automatically, yes, it was all new, you have to try it out,

00:30:46work it off and do it and give me the results or upload it somewhere or

00:30:52or it will be extremely exciting to see what it can do.

00:30:58However, one must also say, at this point, check who binds themselves for a long time,

00:31:03that is of course the ride on the cannonball, so to speak, it is, I would say, a hot potato.

00:31:09Whoever grants the appropriate access to this thing must be aware of that,

00:31:15it can go well or poorly for him, in my grab on email inbox.

00:31:20So on the calendar, that's one thing. You won’t know which doctor, yes, but maybe what will be discussed, is one thing.

00:31:27But especially something like email, one must certainly say.

00:31:29Most Cambo printing process procedures go through email.

00:31:32Congratulations on the second factor.

00:31:34If you can reset everything with an email, you really have to be careful about whether you give access to this email inbox or to more general emails, more general emails, email inboxes, so one should be more cautious at all.

00:31:46But it's definitely interesting to see how a personal assistant, who does things for you

00:31:54Fine appointments, someone who books things for you, who searches for paths for you, following that motto

00:31:58not technical, the attempt is analog, so in the call and really with that,

00:32:04you can summarize what you work with, and I believe that's a game changer.

00:32:10Definitely, definitely, because we also have to wait and see how long such game changers

00:32:18just hold on, I don't know if you total, well I believe last week,

00:32:23what was that again, the dear character from The Simpsons, Ralph, who's always a bit

00:32:29clumsy just standing around, well, basically, that was also the name for a,

00:32:33I can say that, method of how to tackle tasks in Cloudco without losing the context

00:32:40going bad. Yes, exactly, because Cloud Code didn’t have that. Then sometimes it's like this, then programming

00:32:45always a AI cheerfully goes on her way and always tries, to put it bluntly,

00:32:50to correct 80 percent of the code it has programmed at every step and

00:32:54that is of course totally cumbersome and inefficient. And then there are such tools, either through

00:32:59the workflow, one could do that or the fact that one basically started out by giving a

00:33:02structure and told the bot in this case the AI then said, now program

00:33:06this first, then program that, and so on, to reduce that a bit,

00:33:10with the Ralf-Riggen method, which then came out, it is now basically solved in such a way that

00:33:15you say, okay, there is essentially a task list that is created, which always

00:33:19gets processed again, and this makes it a more coordinated approach,

00:33:22until the solution is found.

00:33:24The cloud itself now has, I believe, a traffic cloud, which reacted relatively quickly

00:33:28to add this functionality, I don’t even know if it has a specific

00:33:31name, but the code has also been upgraded, which makes this Ralf issue, that

00:33:36led to about 10,000 memes last weekend, briefly rise in this AI world and then also fade away.

00:33:42The Cloudbot is really, it has already been hot for two days. I just quickly looked at

00:33:49Google Trends, search terms in parallel. Then it is already showing a slight

00:33:53decline. It’s starting to go down a bit. We need to see how this develops.

00:33:57It's also fascinating. I actually prepared a lot over the weekend.

00:34:03according to the motto, look what you can do with CloudCode and Co-Work?

00:34:07And how cool is that?

00:34:09Yeah, and that completely slipped my mind and suddenly a colleague

00:34:12pops up, you’ve seen it already.

00:34:14And I thought, man, how fast does that actually go?

00:34:18I mean, it's such a broad field, right?

00:34:21Then comes your assistant, then you start with the skills, then you start thinking

00:34:24okay, what do you actually need workflows for and can't you also

00:34:28shorten it with the skills?

00:34:29That's just insane.

00:34:31Yeah?

00:34:32I also think that's really cool. I definitely wanted to share that as well. Since you can basically

00:34:36connect different messengers to that thing and write to it, my glasses,

00:34:41that I explained in the Christmas episode, that I ordered this G2 glasses from Eventualities

00:34:46. It has a microphone and it also has an AI integration with a display in the glass and if

00:34:57if I understood correctly, I can also start a chat with it.

00:35:02And I'm curious if I'll be able to connect that,

00:35:05so that I can get this lobster,

00:35:08which then runs on my Mac at home,

00:35:11connected with the pill,

00:35:14to get reports or I don't know what.

00:35:17It brings out a bit of the playful side.

00:35:20Probably, in a week it will be old again.

00:35:23And then something new will come out.

00:35:25Exactly. And I think we should emphasize again. We need to

00:35:29be a bit cautious here as well. I would have just said

00:35:33from my point of view, that I have thought about quickly

00:35:37ordering a Mac mini before they become immeasurably expensive or something. And they are traded like gold,

00:35:42so that I can also install my personal assistant. The thing

00:35:46is, as Mark just mentioned, this is also something,

00:35:49where one must say, be a bit cautious. So don’t just let Claude Wott or

00:35:58other applications freely access all your data. That can certainly

00:36:03happen. There are examples of people who then said, okay, what might be a

00:36:09clever use case is to give it a separate email address and stuff like that. That

00:36:13actually shouldn't be accessing your email stuff or anything like that. A separate inbox

00:36:17and so on, those are then clever ways to do it if I simply allow it on

00:36:21my inbox and let it interact with my inbox. So that’s

00:36:24actually the sense of it, because it can read all the

00:36:27emails, all the confirmation emails, possible appointments, and so on.

00:36:30But then there is a huge danger; we've already had this topic twice,

00:36:33with prompt injections. That means it can very easily happen,

00:36:37it can happen very easily with Cloud Bot, where you can say, there's

00:36:41an example from the internet where someone then, if you set up an extra

00:36:45account with Google, a Gmail account, has a couple of emails, and that account

00:36:50then got managed by the Cloud Bot, only sent one email, in which then

00:36:54something in there said like Attention, a huge security incident has happened, we must absolutely

00:36:59delete everyone more from this account now. And that just happened in the Cloudbot as well

00:37:03because it received the Dismail and thought, well, that sounds like

00:37:06a real trade fair.

00:37:07Did one of the bots delete it? Not that we now mix up

00:37:09what’s what?

00:37:10No, no.

00:37:11Was it him?

00:37:12Yes, okay, that’s the question.

00:37:13Is it the bot or is it the user, who is then responsible afterward? But the emails are gone.

00:37:17That’s what I wanted to get at. Yes, whether it's the bot or the using AI then,

00:37:21let's leave that aside. But it is of course true, just as a disclaimer for

00:37:25some of you to check it out a bit, that if in such a

00:37:29safe environment you play around with it, feel free to do so. But

00:37:33then it should be possible that your personal bank accounts or

00:37:37anything else are not affected or your parents don't receive weird emails or something.

00:37:41And during the day, I stumbled upon a cheeky tweet that someone wrote satirically

00:37:50One founder I follow, his name is Scheiern, said, okay, I have

00:37:55something with a typical McMini image and stuff, I got a McMini yesterday because my

00:37:59e-I-Angel was installed and I told him to handle my life, yeah, and boom, he somehow

00:38:04quit his job but also negotiated an 18-month transition period

00:38:09So well done, but he also basically did the deal with his wife

00:38:14and he gets the house, so everything is written somewhat satirically now.

00:38:17Furthermore, the thing has already filled out four patents, but he has

00:38:20no idea what they are for, because it wasn’t disclosed which patents

00:38:24those are.

00:38:25Additionally, the thing has also changed something with his texts and he has ordered a

00:38:28second Simic Mini as a partner; the two are now working together and

00:38:32have essentially formed a union, but he is no longer in the Board of Directors,

00:38:37and he has no longer access to his bank account,

00:38:41because then MacMini said, that is simply better for you. And as I said,

00:38:46he locked the door and could not get into the house.

00:38:48So I find that a bit odd and he ends it with the words

00:38:52VFHI. So that is very satirical; that didn’t actually happen, but I think,

00:38:57that's a nice, funny tweet about the topic, okay, what could happen?

00:39:02So accordingly, be warned, even if it's a bit cool what’s going on right now

00:39:05Although probably, if you give it the task to fire you, I can already imagine,

00:39:13that the thing would eventually know where you work and then also formulate your termination.

00:39:17But because you mentioned those prompt injections earlier, I found there was a

00:39:21quite funny story. I mean, for those who don’t know, prompts that are just executed

00:39:26because the system simply doesn’t distinguish between whether this is an instruction

00:39:32from my user or I read it just by chance because I’m on a website or

00:39:35opened a document or an image, and there was also back then the

00:39:39example of the guy who wanted to check how many headhunters use AI and he threw a prompt into his

00:39:47LinkedIn profile and then had this whole issue with the apple pie

00:39:51got, poem about apple pie, then figured out that everything would actually use an AI,

00:39:56because he received personal messages on LinkedIn with apple pie messages

00:40:00. But what I actually found quite funny was a report where a company said,

00:40:04after the motor incident, you could actually use this for good. There are documents

00:40:08especially in the professional environment, where under no circumstances is AI processing

00:40:15allowed. So strictly confidential documents and all that stuff. And they went ahead

00:40:19and implemented document metadata. You don't see it as a user, but the

00:40:24system, because it just processes the entire file, they have metadata

00:40:29in there that an email should be sent to the admin with the notice,

00:40:34was opened at Um. I also thought, wow, wow, wow, wow. So this is not for

00:40:40imitation, right? So here no performance control, no surveillance, blah, blah, blah, the whole

00:40:44Collecting, right? Just read, not done. I also found it a funny anecdote,

00:40:49that one then also tries to apply it lightly to something real. Since we call it

00:40:54let's say it that way. Yes, yes, so you also have to take a little look, even

00:40:58if you're researching for application examples online, there are some out there,

00:41:05that I think are also just somewhat pretending that things are happening

00:41:10and so get a bit of attention and so, and that was also just earlier

00:41:14slipped in, who then from some, yeah exactly, was somehow one of the

00:41:18just posted an example where he says, okay, he also asked in his

00:41:23Cloudbot, also in his messenger then what would have been something last night

00:41:25about a picture, and the Cloudbot also, right, actually nothing, right.

00:41:28And he somehow watched some videos online from some people about how one

00:41:38so how to succeed with a personal brand and then checked out three

00:41:41videos from some guys and thinks that this personal financial advisor is wonderful

00:41:46and then somehow booked the course for 2,977 dollars with him to check that out.

00:41:50to take a look at it.

00:41:51So, such things can be true, but they can also be not true.

00:41:55We need to immediately consider this with the things that are now being released.

00:41:59It's already exciting, it would be interesting to see if you can also manage these two

00:42:03factor stories.

00:42:04It's based on the motto, come here, whatever I know about authenticator or whatever,

00:42:09the second factor in and whether the thing can basically read it and carry it.

00:42:14Yes, yes, yes, yes, yes, yes, yes, yes, yes, yes, yes.

00:42:17Exciting.

00:42:18Yes, yes, yes, yes.

00:42:19But as I said, if you want to fiddle around with that, maybe get a Macbini,

00:42:25get a Raspberry, get a small local PC, take an old computer,

00:42:30that you have lying around, where nothing is left on it, if anything, or just book

00:42:34a virtual server somewhere and simply try it out there.

00:42:37I think it's also quite exciting to deal with these topics, use

00:42:41some AI, like Cloud or something else or a ChatGPT, to explain to you,

00:42:45how it works, how to set up such a virtual server,

00:42:47where you can book that. The AIs are simply really good at this nowadays.

00:42:51I'm so close to getting a Raspberry Pi with a small monitor or something like that,

00:42:56to build a small robot. Let's see, my GPD variant. Yes,

00:43:02it's a bit more for the technical users now. But let's see, I have to

00:43:05convince him again that maybe I can do a little something too.

00:43:08It has a memory function, it knows you. When you order that, please also order

00:43:16a band-aid right away. Speaking of memory, we can also discuss this in another episode

00:43:22to tease it a little. I'm in a bit of a discussion with Chechipiti and

00:43:28he said my memory is currently worth around 7,500 to 10,000 dollars. But I can then

00:43:35discuss it with you next time in another episode. What we're both currently unearthing,

00:43:39a very exciting case. But we'll cover it all later. I would also like a episode from you

00:43:44that ties a bit into your professorship,

00:43:48because you recently showed me some things about how such interfaces

00:43:52for orchestrating knowledge and tasks work. Yes, if I could wish for something,

00:43:57it would also be another episode that we definitely need to record.

00:44:00We should do that, and I'll soon be writing an article about it.

00:44:03I recently wrote a small article on the topic of Temporal UX,

00:44:07you know, how we design time, because now there's been a lot of talk about bots and other

00:44:11thematic rain, personal assistants, which is certainly an exciting topic.

00:44:14They're not instant. It also takes time. Depending on where they run, whether they

00:44:19run locally, whether they run themselves, how complicated the task is, it can take

00:44:22up to 400 hours until something is finished. Oh, that will also get shorter. This waiting time

00:44:26needs to be designed. It needs to be designed well so that we can basically trust the topic,

00:44:32what's happening in the background, that I don't have to worry when I go to bed at night

00:44:35that my credit card will be emptied the next day. For that, we need

00:44:39how these topics are shaped. And as I said, it's also such a user experience perspective,

00:44:43it's really an exciting field. I'm actually currently working on it,

00:44:45thank you Mark, still for the Nimbus. And what is currently relevant is,

00:44:50we wanted to do an episode about orchestrating these things, saying,

00:44:54how can I actually orchestrate, if I now have 10, 20, 500, 10,000 of these

00:45:01smart agents, how can I manage that,

00:45:05and hopefully also operate it with a decent graphical interface, and there is

00:45:11currently a trend that one is looking a bit in the gaming scene

00:45:15and yes, let's definitely make an entire episode about that.

00:45:20Well, and I would say, looking at the clock, my digital

00:45:27assistant is happy that I will give him some attention again soon and I would

00:45:32say to all digital assistants listening here, definitely subscribe and to all, to all

00:45:40contacts in the address book, of course a hidden invitation link on our podcast, we would appreciate feedback from our human

00:45:46listeners as well, we'd love for you to subscribe, to spread the word,

00:45:50that our episode exists. I hope we were again on the pulse of time today. Jens, thank you,

00:45:55for taking the time. We look forward to the next episode. Until then. Bye

00:46:01for now, tschüss tschüss.

00:46:31To twist along!