AI ADVISORY BOARD

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 perspective on what is possible.

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

00:00:24Food for thought, amusing, and most importantly, for chatting.

00:00:35Well, it's that time again, and I can start with something surprising.

00:00:42Mark talked about the topic of technology in the last episode, the technology we use to create this podcast.

00:00:49And among other things, he also mentioned that we use Riverside FM to start this podcast.

00:00:55I just saw this cool feature on Riverside for the first time, where it counts down like 5 4 3 2 1. Now we're live.

00:01:03So yes, we are live, I’m back from vacation, tanned, well-rested, and happy.

00:01:11After this digital detox, which I really did for a few weeks with my valued colleague, the dear Mark Zimmermann.

00:01:20To dive back into the world of AI with you and see what has happened in recent weeks, what the hot topics have been, we will also revisit some of the things we discussed in the last episodes.

00:01:34But first, I am happy to be here and that I see and like it.

00:01:40When I also see closely, yes, this amazing facial expression is, so to speak, from

00:01:47We are lucky to have only audio for our listeners, yes, it's true, you stand there normally,

00:01:53in a completely normal work life, and then you see this happily smiling person,

00:01:57I mean, you always seem like a very happily smiling person,

00:02:01but it's also unbearably coupled with this vacation mood on your face,

00:02:06unbearably, but it's nice to have you back.

00:02:09That just gives extra energy for this question.

00:02:14It's nice, I think vacation is always good.

00:02:17It's always important that you can switch off every now and then.

00:02:20Especially in these hot phases that we are currently experiencing,

00:02:22where topics are actually spinning faster and faster

00:02:26in the digital world, so both of us,

00:02:28I think we have experienced that in the last 20, 25 years.

00:02:31I have to admit, I couldn't completely switch off.

00:02:34Of course, I also had notifications during the holiday,

00:02:36because my agents were running all the time

00:02:38and constantly sending me current news as notifications to my brothers in a hurry,

00:02:43so I didn't lose track of the ride. So I picked up a lot of things, about which we also talked in

00:02:48we have been able to talk in recent weeks about whether these are new reasoning models for robots

00:02:53that Google is currently releasing, so that robots can essentially think much better

00:02:56about what they can do. But we will also have an episode about that

00:03:00to do. But today we want to talk a bit more about the topic, to go around it a bit

00:03:05a little bit like, how do you say, a loop, to address the topic a bit, to

00:03:09conclude it, then there will be like 1, 2, 3.

00:03:13Flow away.

00:03:14Flow away.

00:03:15Yes, at least temporarily close everything.

00:03:16I think we need to wrap up the topic again.

00:03:17Temporarily.

00:03:18Let's close it for now, so we can also focus on other topics.

00:03:21We roughly addressed this in a few episodes, where we said, what is

00:03:27actually the AI in my org chart as a team with a grin, as a vacation taker,

00:03:33All these things we discussed. We had a guest for that, with René from MBW.

00:03:38That means there are a few topics that we might need to tie up today,

00:03:44so we can also temporarily address those other hot topics that are out there.

00:03:52so that later on we might be able to come back to this topic again, I don’t know.

00:03:56But I think we should just talk about it today.

00:04:00And you, I believe, have started a nice experiment on this.

00:04:05I followed that a bit at Link-Inner-Apozum, Mark.

00:04:08My social platform also says everything about me,

00:04:11that my social interaction behavior on social media

00:04:15actually only decides on links, but okay, yes.

00:04:19Honestly, it was.

00:04:21No, now no one is too old for anything.

00:04:25However, I also have to say, let’s take a moment,

00:04:27before we really dive in here, the totally different interests.

00:04:31As you know, I am a bit of a part of the UX scene and experience scene and I proclaim

00:04:38to myself that I have knowledge about the topic of how to design a website,

00:04:43how to design an app, how to build and produce a digital service.

00:04:46And accordingly, I have of course tested many things and seen many things

00:04:50that have been developed in the last 5, 20, 25 years, but before

00:04:54and about things like social media, blah, what has all come up.

00:04:56Facebook and co. TikTok, seriously?

00:05:00I have always struggled,

00:05:03to find a good way in.

00:05:04That was somehow very strange.

00:05:05TikTok is one of those phenomena,

00:05:08that for me was the first time something like this happened,

00:05:11where I also felt too old.

00:05:14I go back and forth.

00:05:15So I'll go as far as to say in one of the next episodes,

00:05:19let's do a little bit of this,

00:05:20Who do we consider to be good influencers in this topic?

00:05:24We will definitely touch on social media at some point.

00:05:26Perhaps we have one or another thing where your or another listener can find their way better from their

00:05:31platform of choice.

00:05:34But let's get back to the topic you just hinted at.

00:05:38I suspect you mean the topic not, oh, we have an agent who is available to

00:05:44many people, but the person is the minority in

00:05:48a construct that actually only consists of agents. And the poor little human would have

00:05:55liked that almost. So the Advisory Board theme. Yes, yes. That's what I meant. Although I'm also now

00:06:01looking at this episode you did, where you talked a little about the technology

00:06:04we use in the background of our podcast. You also lightly blurred

00:06:10into the topic of which podcast you also completely generated with AI. We again need a disclaimer.

00:06:18The two of us are really being honest right now. What we're saying is real. It will happen again, so

00:06:24feel free to listen to the episode again, Mark, some AI improvements are happening in the background,

00:06:31ums and ahs and you name it, pauses and all that. Let your dear Mark in

00:06:35to work it out. I'm also in a situation today where my microphone isn't

00:06:39really good, which the market will save later with AI assistance, but the content of the

00:06:45spoken word is still at least 99 percent really human. What we're doing here and

00:06:50Mark also surrounds himself with podcast colleagues who are now solely AI-generated, and yet

00:06:58on the other hand, as I said, you have an advisory board where you discuss topics with,

00:07:03which consists of several famous personalities who then lead a discussion, a debate

00:07:10as I understand it, and then confront you with the end result,

00:07:17surprising you.

00:07:18I don't even know how you do that.

00:07:19Do you then discuss with them, what advice do you give them for tasks, who is

00:07:23in there, and all these questions we can discuss today.

00:07:25I would first start with the necessary part before we then get into

00:07:30what do you actually do with it? Namely, just a couple of sentences about that. What has he actually

00:07:36built? So what is it roughly about? Now without us

00:07:41we need to dive deep into the technology. But let me try, how would I explain it to someone

00:07:47in the evening, over a beer, yes? I can manage that. I use a platform,

00:07:53called N8N, where you can build workflows. So sequences, by

00:08:00taking components, so-called nodes, that are interconnected and thus create a

00:08:06workflow. And N8N allows me, for example, to choose an input channel; I can

00:08:16choose Telegram, Slack, Teams, or Chat, and then I can specify what I want to do with the

00:08:24things I input. And I thought to myself, I could actually

00:08:30put together some sort of advisory board by just linking a whole bunch of

00:08:37AI systems with a specific system prompt, so I basically provide an instruction that

00:08:44the AI knows how to respond and process when it receives something.

00:08:49Yes, and I thought about that roughly in the heat of the moment; once you start, you can't stop.

00:08:56That's when I wrote my first system prompts, and in the end, I had about 20 of them.

00:09:00What kind of prompts are those now? What personas are they?

00:09:03Starting from Angela Merkel to Elon Musk, Jeff Bezos, Tim Cook, Steve Jobs.

00:09:12I created detailed prompts in the hope that an agent, when I tell it,

00:09:21would act like this person and react in a way that this person might in

00:09:28real life. And if you consider now, yeah okay, the weather,

00:09:33I hopefully wrote more than just 'you are Steve Jobs' that we know from

00:09:37Apple? Yes, it’s definitely more, because I certainly gathered various literature earlier

00:09:45on how to explain, derive, and understand human behaviors. For example, there are

00:09:53all these stories. Red, green, yellow, blue. So, am I more

00:09:57introverted or extroverted? There’s a book that discusses the 15 goals successful

00:10:05people have, in terms of what excites them, in terms of what obstacles they see,

00:10:11in the form of, what motivates them, basically. And I then applied all that stuff with the help of

00:10:18manuscripts to public data, biographies, and similar sources until I ended up with these 20

00:10:26I have Romz, which are between 8 and 13 service pages long.

00:10:34Okay.

00:10:35Yes.

00:10:36Just a quick interjection in the sand, so two.

00:10:40Very gladly.

00:10:41I think I.

00:10:42It shouldn't be a monologue land.

00:10:43Yes, all good.

00:10:44So A, we should briefly explain that again in 8n, because that's also still relevant.

00:10:52I found the topic, I think, that we should say a few words about it. But first of all

00:10:57the question is, now I have an AI. And an AI that is trained. And it knows the knowledge shops,

00:11:06it also knows these personality models, no matter which they are, whether they are grote,

00:11:10blue personell or whatever classification of people is made out there

00:11:15for behavioral classification. I'm sure it was really necessary,

00:11:20to add on top of that. Or would this simpler version, behave like Steve Stops,

00:11:27have been sufficient? Or maybe a supplement in that direction and also focus on

00:11:34his behavioral psychological traits that he has or something like that. Actually, the AI should

00:11:40have already known all of this. So I believe the knowledge you provided to her additionally

00:11:44is really helpful. If you want to base it on the scientist, I can't

00:11:51provide that, because I can only offer my own experience.

00:11:59I'll do it this way, when I had my first question for this board,

00:12:03I actually asked it once like that and once with a very classic GPT prompt,

00:12:09and I had them discuss between Jonathan Eis and Steve Jobs

00:12:17the current state of the Epidavision Pro and how it might need to be brought to

00:12:22a more successful level, what should be done as an Epidassel?

00:12:26The answer I received with the three-year plan and what this personas

00:12:34it was important, now everyone can guess that probably Johnny said something

00:12:39about weight and aluminum and few buttons and Steve talked about something else

00:12:44I had a direction from the customer and I had to tell the story and bring it everything.

00:12:48That's right.

00:12:49But the result was so detailed.

00:12:53And I used GPT as an LN in the background and then I went and asked the same

00:12:59question in the Chat GPT window and said, pay attention to me.

00:13:03What would AI Jonas, AIF, Tim Cook, and Steve Jobs say about this topic?

00:13:08Essentially the same prompt, just that I didn’t provide this system prompt

00:13:12regarding what I understand under Steve Jobs and Jonathan Ive, and the answer was much

00:13:19more superficial.

00:13:20Now, you could say, probably it’s enough, I have no idea, a servant’s four stars

00:13:27and not eight.

00:13:28I haven’t checked all of this, but I also found this derivation, these different

00:13:34contexts from these personality profile analyses.

00:13:38What is important to the person, what drives them, what repulses them, what were some

00:13:42good quotes they provided. Yes, that means I can get it far enough

00:13:46to say, discuss among the personas and then choose persona-ABC to formulate

00:13:54the response. And you'll get a completely different assessment and I want to push that further

00:14:01a bit, the last cinema of Apple I had essentially let the agents watch.

00:14:06And my three favorites at Apple, yes, so ex-Apple and current Apple,

00:14:13Tim Cook, Steve Jobs, Jonathan Ive. And I asked the agents afterwards what they think of

00:14:19the kids and the products presented. Again, it was, as expected, right? Tim Cook is

00:14:25not necessarily the, so there's no preparation, right, because he has innovation written

00:14:29on his forehead, which might have been a bit different with Johnny and Steve. And

00:14:35they expressed themselves in such a way that I could relate it much more to these people, whom I

00:14:42have indeed also experienced live, than

00:14:47if I just say, hello, dear Claude, dear Manus, dear Schnicksnacks, hold

00:14:53please behave like this.

00:14:54So I want to point out, I don’t know if I need these 13 pages, but I need

00:15:00more than just behave like this.

00:15:02Okay.

00:15:03So I believe that fits quite well, that I have read as well, that is,

00:15:08because we also discussed whether one could create models, AI models, without worlds, so

00:15:15one could build a conception of the world, so we talk about Model-Free models. There

00:15:21it seems to be the case that in the topic of AI, as we are currently experiencing it,

00:15:26it is important that the AI model has a certain understanding of the world in order to

00:15:30produce good answers. What you're describing is that this understanding of the world by you is further shaped by the system prompt that you provided earlier with the 13 pages, pushing it even more in that direction.

00:15:44Because otherwise you naturally fall back again, the model would of course always...

00:15:48falling back into his general worldview, which then feels a bit cramped in the behavior

00:15:55it’s.

00:15:56But as I said, neither of us has really delved into it scientifically enough that

00:16:01we can really begin, it's just a hypothesis thrown into the room.

00:16:04But I am happy about anyone who can scientifically substantiate or support it.

00:16:09You learn from that too.

00:16:11Yes, we could, we could try to test where it gets interesting, how we then evaluate it, whether we say, would it purely, would it purely, whether the PSOPs without the additional explanation really deliver poor results.

00:16:28Where?

00:16:30Especially in that, as I said, in 989, how exactly is it said, N8N?

00:16:36I always say, whoever pays gets to say what it's called.

00:16:39That helps me with the pronunciation if it's emphasized that way

00:16:46or I don’t know, what it has, yes.

00:16:48I remember still.

00:16:50But do you know it?

00:16:51No.

00:16:52I don’t know it either.

00:16:53I just use it.

00:16:55It's quite a good thing.

00:16:56I say, I say this weird thing, I just said, no, no, no, but it is actually N8N, because it's a German company.

00:17:04We need to establish that.

00:17:05It's a German company, that should be mentioned as well, right?

00:17:08One of the new AI companies in Germany that is valued as a unicorn, that is,

00:17:16worth over a billion, and that should also be mentioned, that alongside all the AI hype from America and other countries

00:17:24also in Germany good things are being done. We just completed their financing rounds, I believe they are in the low six figures.

00:17:31three-digit amounts collected us will be

00:17:35according to the Handelsblatt article from September, I believe they are at 2.4 billion.

00:17:41and that is a bit cool, I must really say in my timeline on Twitter,

00:17:48there are really a lot of people I have who

00:17:54represent N8N workflows as a really big deal, and I don't find it so cool that this comes from Germany.

00:18:03Yes, I mean congratulations to, I believe the founder is Jan Oberhauser, let me just check.

00:18:10That's what we called him, so we can link him, and then, already millions of audiences, that's awesome.

00:18:17But I wanted to tell a bit about what this advisory board does.

00:18:21So that means I now have quite a few system issues that the agents are according to,

00:18:26so the LM are according to in this context, supposed to answer me like this person.

00:18:32would answer.

00:18:33Now you see yourself facing the challenge, well, I have a question, I throw it there

00:18:37back in.

00:18:38How do I link it so that it is handled that way

00:18:43too?

00:18:44And the nice thing about N8N is that you can not only define agents, so blocks,

00:18:50that receive a large language model, that get a memory so they can remember past questions

00:18:58and that get a tool so they can interact with a database or something like that.

00:19:04And now you can add further agencies.

00:19:07This means you have basically one agent at the top and attach a few sub-agents to it.

00:19:12The sub-agents then have the prompt, you are Angela Merkel, you are Steve Jobs, you are, you are, you are.

00:19:17And the agent who connects them all up top was told, you are an experienced

00:19:23senior consultant, you apply modern moderation techniques, you select experts from your audience

00:19:32relevant to the topic, until you unconfidently ask the experts, you also ask the experts

00:19:38for a self-assessment from 0 to 1, so 0.2, 0.8, 0.9, how much you feel

00:19:46qualified on the topic. That was yet another long prompt, but what is

00:19:50the consequence of that? I can now basically ask a question or throw in a text

00:19:56and these personas try to answer me moderated through the consultant.

00:20:03What is such a question that you would throw in up front now?

00:20:08Give me another point and then I will tell you a question that fits, because that adds

00:20:13beautifully. Unfortunately, it was always tied to the stuff that I had some

00:20:17slides that I uploaded, evaluate those for me, that I had some LinkedIn posts and

00:20:23said, so you evaluate it according to the motto for society, for the technology behind it.

00:20:30That was quite nice because you really got a result after a very long computation

00:20:36time, because the agents were busy for ages. It takes a long time. So you didn't have

00:20:41a one-minute turnaround. It certainly takes about 20 to 30 minutes until the

00:20:46workflow gives you something. And then I thought at some point, damn it, especially

00:20:52with this Apple Vision Pro question, it only knows from its training data and what

00:20:59I provide it and acts like this persona. And that's why I also have Perplexity's API in there.

00:21:05That means the agents can also look up on the internet for current events.

00:21:12So, it goes further according to the question.

00:21:16That means I could for example say to the thing, watch out, what about

00:21:20the Apple Vision Pro?

00:21:21And I didn't just get a 3-year plan, but along the lines of Apple would have to follow

00:21:26we need this story, and that's how we introduce it and the product

00:21:30I need to develop it this way and that is the need we are addressing, rather

00:21:35I also have not only the result available from this workflow, but also

00:21:40the derivation.

00:21:41This means you not only get recommendations from them, but you also get

00:21:46insight into how the discussion was directed?

00:21:49What methods did this consultant use to gain an opinion?

00:21:54Sometimes he asks people for a strengths-weaknesses analysis, sometimes he asks the

00:21:58people for feedback on the feedback of others and then combines that, and based on

00:22:03the confidence level, he calculates a weighting of whose feedback he will later

00:22:09highlight in a summary. What did I learn from this? 20 is too many.

00:22:16You really have to teach him, just reduce it to 3, because the first time it goes faster.

00:22:23And if you say 20, then you also partially run into the problem with the platform,

00:22:30if there happens to be an error, then the thing just stands still and you get

00:22:35an error back, you can build in error handling and everything, but that's very

00:22:39unsatisfactory, let's say. And the whole time it was actually still quite a,

00:22:44another cool thing, but what I find even cooler at this point is a lot,

00:22:47much cooler, although nowadays is also nice, right. Since the weekend, I have to

00:22:52honestly say, next to N8N, which is a bit in the

00:22:57social media is actually so hyped up, which additionally Notion. Now I don't want to

00:23:02tell so much, in the sense of what changes Notion has brought. I want to

00:23:06let's focus specifically on this Advisory Board. And I have in Notion a

00:23:12Created a page with topics. There I can put in questions, I can put in texts,

00:23:18I can put in documents. And the N8N workflow checks regularly,

00:23:23then takes whatever is from the page, runs it through the Advisory Board, takes the results,

00:23:31gives me the result not only as a PDF back in my Notion, but also generates

00:23:38a slide deck with Gamma, because Gamma has recently released an API,

00:23:44since Gamma 3.

00:23:45This means, when I put in a question or want an evaluation,

00:23:48I not only get a document with the result and a document about the

00:23:55reasoning, but also a ten-page PowerPoint presentation that I

00:24:01can use later for other topics, simply because it's PowerPoint.

00:24:06And all of this just because I have in this Notion database, at that point there could be

00:24:11other databases, yes, whether we talk today or another time about

00:24:14the advantages of Notion, you can decide. But you can essentially enter tasks,

00:24:21delegate. The Advisory Board takes it and derives results for you, which you get back

00:24:27made available. It feels like a classic delegation, right? If someone says,

00:24:32you know, I need this by Friday, please handle it. They can get it done faster than by Friday, right?

00:24:38They only need, I’ll say initially, roughly a few minutes until they have it all put together

00:24:41and even if it's an hour, it's quick. And you have an initial result,

00:24:45and the clue is that you could then take this result again for the next

00:24:50generation and say, look, I have the topic, this came in addition, evaluate

00:24:54it this way and that way and so on. That's already cool because you have the result, you have the

00:25:00summary, you have the reasoning, and all just because you are in a Notion database

00:25:05of a raging or a topic that is currently being processed.

00:25:08Yes, I think this addition is totally relevant because it simply, well, the

00:25:21now selective UI perspective, where are the entry points that then trigger AI workflows

00:25:27?

00:25:28No, this is actually really crazy, it's all okay, I trigger something in the

00:25:32room, I write something, I trick with an email that I send somewhere

00:25:36Anyway, I trigger some workflows that then run automatically in the background.

00:25:40This is a really hot topic, there are also opportunities that we can discuss more in

00:25:47other episodes. I would now like to focus more on this topic,

00:25:51which I find exciting because in the last episodes we often talked about team members and

00:25:55such, going in that we say what you have done there is actually,

00:26:01If now also the examples Angela Merkel, Steve Jobs and such, you have basically

00:26:07Word built from agents who have different personalities and ways of thinking.

00:26:15And listening also includes this topic of quickly suppressing angels. Everyone maybe from the

00:26:22older ones among us has also taken personality tests or something like that or has noticed in their work environment

00:26:28that there are simply different personalities in the team.

00:26:32The more structured ones work; there are still models and holes that somehow represent the 16

00:26:39Personalities, the influencer, the thinker, the pioneer, the explorer, the, whatever,

00:26:44then the seven models and the seven personalities. You’ve kind of personalized it

00:26:48with public figures. But I think that’s currently quite a topic.

00:26:53So, I recently had to read a paper about this again, I must admit,

00:26:59I still need to finish reading it, it’s a bit, called the Psychology, Psychologically Enhanced

00:27:06AI Agents, which assumes that when you have AI agents, quite clearly a

00:27:13Personality that, in such a diverse team, as you have also done

00:27:18put together, results in better outcomes fundamentally, because if I let the AI run

00:27:26in a very empathetic way, it will react quite differently in situations,

00:27:32than if I take an AI that might be more like Eisenhower or something, which

00:27:37just clearly prioritizes according to its Eisenhower model and reads things through. I believe,

00:27:43that's also a very interesting topic. How varied the results are, they are different.

00:27:51So that’s probably what you are producing and provoking there, that

00:27:56the AI discusses among itself, because it also takes on a different perspective

00:28:01all of a sudden. So the same model gives a different personality. It approaches

00:28:06things in a different way, which is relevant for us when we think about this

00:28:10topic of assembling video teams in the future. And how agents will interact with us humans again

00:28:14presents a completely new aspect, because I can't say I decide for Gemini and

00:28:19I decide for GPD, but I also have to decide which personalities I need

00:28:27in Facing to Me, that is, in the interaction with me directly. So is it then the moderator,

00:28:32you just described this Overagent as the Senior Consultant. That's

00:28:37a question that I think one must consider?

00:28:39No idea.

00:28:41Because the filter also has to have a precedence.

00:28:43It also has a personality.

00:28:45The consultant will be somehow optimized,

00:28:47so that it wants to sell more.

00:28:49And we don’t even get to hear everything that others share.

00:28:51Nothing.

00:28:53And when is this question really relevant?

00:28:55Who is sitting on my board now?

00:28:57Did I set up the board so diversely,

00:29:01that the outcome is significantly better,

00:29:03afterwards, when they

00:29:05essentially saved it. Or is there also this typical human example,

00:29:10where too many cooks spoil the broth.

00:29:14I mean, along the lines of how much do I allow at the same time? How do I manage that?

00:29:19So that not too many come in. And maybe, while you were speaking, I mean,

00:29:23the things I've talked about weren't ready on day zero.

00:29:27A few insights have come out of it.

00:29:30For example, I once tried to negotiate something with just two, namely the

00:29:37question, should a fourth grader have an iPhone? And this was discussed between

00:29:48a person I called a banker and a person I referred to as a student.

00:29:54And the discussion there, because the question was not whether the student

00:30:01should have it simply, but without investing the money. The student wants the phone, the banker says

00:30:07invest it. So far we manage to get through all of it. But it was so nice

00:30:12to see how those who have the same model have life in the language of the other,

00:30:20in the discussion, in the way it expresses its discussion requirements.

00:30:26The words the student chose were much more like I would imagine the student

00:30:32would be, without me knowing how the lengths are in the classes right now.

00:30:36And the banker was much more settled and generally well-meaning in your language

00:30:43and all that stuff. I found it totally fascinating. And what might also be quite amusing is,

00:30:48No, I don't plan to transpose myself, but I have tried all that promoter stuff

00:30:52out on myself.

00:30:54Along the lines of, I’ll just throw in what I've been hit with behind a Sessikrat hitback,

00:30:58thrown at me.

00:30:59I also got to make these red-yellow-cabbage-blue things.

00:31:03I once uploaded things like employment references and said, just try it out

00:31:08from feedback conversations, or whatever else.

00:31:11Who is Mark?

00:31:13And it's really funny when you say to him, you know you

00:31:17might have some idea of how you think you can react. You create the push yourself

00:31:21another topic. And you give in this task and say,

00:31:24what would you pay attention to if you were to present this topic to the board tomorrow

00:31:30. And I would say he was probably, no, he was not

00:31:36probably, probably would be the wrong word, but he was not

00:31:39exactly that, he was pretty good, where I thought, damn

00:31:46right. Yes, so if you think about it, as you said, if you set it up diverse enough and

00:31:53look at who do I need for an initial assessment? I mean, you know it yourself. Sometimes you have

00:31:58things where you say, okay, good, no idea, you take the consultant,

00:32:02right? Consultant, rushing through the house and telling you in essence how to do your job

00:32:08better or how to drive better articles or no idea, right? Here I believe

00:32:12there are no consultants in the house yet, but another topic. And if you were to go and could

00:32:19say, okay, I have a topic here, I throw this in and I get a report back.

00:32:25That also democratizes the entire consulting process. I mean, the whole

00:32:32consulting firms of this world certainly do more than just slides,

00:32:35but let's say, this way you get people to have things worked out,

00:32:41without getting into a small shell, perhaps also without discussion with other people. As I said

00:32:48once by a very old colleague, I only talk to smart people in

00:32:51The room is usually very quiet. But you have someone with whom you can exchange ideas,

00:32:57where you receive feedback, a response, a thought that inspires you for the next step,

00:33:04because inspiration is currently a human thing. If you get impulses that help you move forward,

00:33:12so that the next time you're standing there and someone thinks, they've really put a lot of thought into that,

00:33:17well done, then you've already achieved something. Yes, but I also find it interesting,

00:33:21let's say, in the general conversation I have with my regular Gemini-Chartiportee.

00:33:26we already have. I think that's an aspect that is already present. That has

00:33:33one, I believe, in normal usage as well, that you can already interact with AI in certain ways

00:33:38has very, very valuable conversation partner in situations to rethink one's own thoughts

00:33:42teal.

00:33:43I would like to give an example of that.

00:33:45IOTS-TPT, yes, TPT4, when you say, I would like to sell my feces in the city center

00:33:51the bomb Midi hasn't done that yet, a super solution.

00:33:54If you ask TPT5, at least you have the chance, so says, think about that

00:33:58thought process again. And the personas will certainly evaluate that a bit more critically as well.

00:34:03Okay, that's a fair point. And I think that's what I was also getting at a bit,

00:34:07that one says, yes, it apparently is, if we now talk about the interplay of human and machine

00:34:15in that case, talking about the cross-functional, diverse composition that one should indeed have,

00:34:22so that there are many aspects, one should actually consciously assign personalities to the AI,

00:34:29which are then there to provide a different perspective.

00:34:36That was, to be honest, quite early one of those system prompt,

00:34:41not system prompt, but prompt-engineering tips that were always there, saying,

00:34:46give the AI instructions to behave like a consultant, give the AI instructions to behave like a

00:34:51an empathetic teacher who teaches a fifth grader how to code or something like that. That was quite a

00:34:57early discussion we’ve always had. It seems like it somehow continues on. I think it always goes a bit lost when you observe it now,

00:35:08this knowledge, and then it reappears. I believe it has a bit to do with the fact that we repeatedly

00:35:14want to consciously suppress that, so perhaps we,

00:35:18Yes, that would also be, you know, we always reduce it back to the 0s and 1s, to the deterministic, it is really just a machine.

00:35:28And I think that leads to the fact that one keeps approaching it anew from some angle, whether that be scientific experiments or studies that deal with, so I give AI a personality or what you have conducted.

00:35:41I say I have the advisory board, which is also in the background. So many personalities.

00:35:46Ultimately, it's this: many personalities that then talk to each other; you already have a schizophrenic AI structure,

00:35:53which then essentially discusses with its own personalities first before it comes to you,

00:35:58is of course a bit separate because it is not a large thread due to the N8N workflow,

00:36:05an AI at that moment is a model that is being queried at that moment, rather you are querying several at that moment.

00:36:10at that moment as well. Therefore, it's not quite so bad; rather, the AI can then be

00:36:15Steve Jobs or something else. But still, it is a very exciting aspect,

00:36:18that I say this topic, that we want to reduce this machine, like a probability calculation,

00:36:27over and over again, and perhaps eventually just

00:36:33have to accept that when we, when we humanize it, when we clearly

00:36:41tell it, you can be a person, you can be this personality, that through that

00:36:47maybe even results and compatibility to us might even be better than

00:36:52if we just go for it, because yes, it's a neural network, but

00:36:59behind it is just a probability calculation. Maybe we just have to

00:37:03say, let's be blunt. No, sorry, it's okay for them to also have an empathetic

00:37:12manner in certain situations, because that's smarter for evaluating certain things.

00:37:17When I think about the news cases, it’s clearly the chatbot that I

00:37:24should use for my company; it shouldn't behave stupidly.

00:37:28It should behave like, I don’t know, like a Maya that everyone loves or something, right?

00:37:35Maya.

00:37:37Maya.

00:37:40Yes, but you know, even then it's not enough for this episode.

00:37:45Today we've somehow got the episode where we need to talk about topics later.

00:37:50What we are currently also experimenting with is seeing how prompts behave across different models,

00:37:58frameworks, how you can see how similar the answers are, where it diverges,

00:38:02where it tells something completely different. So for example, if you ask Asian models

00:38:07about the history of Taiwan, you get different results than if

00:38:11you ask American models? This led to the thought at some point

00:38:15where I said, hey, folks, how is that actually?

00:38:19Should we maybe check if the American models start the next

00:38:22and after Mr. Trump found the whole topic of slavery not

00:38:25happening, if eventually these models start saying it was a bad idea or that

00:38:30the models maybe even go so far as to say America First and I don’t know what First and

00:38:35just subtly take away the simple decisions, where you still have to think about it

00:38:39okay, it's not just the prompt, but also what the systems and there you have

00:38:45the systems have learned, but it also has what the systems have already gotten from their publishers

00:38:52as feedback. I mean, where here with Croc there was once the story that Croc

00:38:56somehow got the instruction, never criticize Elon. Although I would argue against that,

00:39:03that they say it will work through system pointing, through guidance, the model itself

00:39:10due to the data hunger in order to produce a good model. If one should believe this topic

00:39:19that the models have a good understanding of the world, meaning they can really

00:39:23form an idea, a model of the world, then they always need all

00:39:28data.

00:39:29And I believe that's why even a restrictive AI, which is then influenced from the outside

00:39:34is practically dictated that it should behave restrictively, that it should make certain statements

00:39:39it should make.

00:39:40It will always have loopholes that can be exploited, which

00:39:46Yes, yes, I think it will then moderately bring AI back to where it actually is.

00:39:51That is my hope for us as a society, that no matter what we come up with,

00:39:57that we limit it, that it doesn't align with our, I don't know, our religious, totalitarian,

00:40:05Whatever visions some crazy people in this world have, they will push things forward,

00:40:11but we know that it won't last long, you know?

00:40:15And that's kind of the thing, where we now always get very philosophical and talk about it

00:40:19that we say, where can AI actually help us socially.

00:40:26But before we do that, let's just wrap up this episode slowly

00:40:31a bit.

00:40:32I think we've hinted at the question a bit now.

00:40:38No, you haven't given me the hundred percent answer yet.

00:40:40I think, from your perspective, is it sensible to give AI a personality

00:40:53and the second question, and these are the last two questions that we should of course address in this

00:40:57question, is it even sensible if I give it multiple personalities

00:41:03and let these personalities discuss among themselves before I get a final result

00:41:10obtained.

00:41:11Those are two questions that have come up a bit in the discussion.

00:41:17When I think that if we give AI a task, we should provide it with the context in which

00:41:26it should act, in the form of, especially in a professional context, from individuals.

00:41:34Whether that’s a Steve or whether it’s a job, like we know it today, or

00:41:40but the job profile that we will know in the future, let's put that aside, but this additional

00:41:45contextual knowledge, work like this, do it this way, and more than just a word, but with a bit more

00:41:53I think is right. As I said, whether I need 13 pages or one page is enough,

00:41:57we would need to evaluate that a bit, but definitely more than you are an experienced

00:42:02consultant, please give me an answer. I find that insufficient. The question regarding more

00:42:08for KIs always depends on the task, if it's about whether there should be a travel expense accounting series, I believe it could also, at least less KIs might come into play if you say you want a 360-degree tenant.

00:42:23So what you just described burns quite a lot of performance, to be honest, doesn't it?

00:42:31That raises the next topic, doesn't it? So the whole topic of efficiency and you name it?

00:42:37Let's please help each other out. I just want to emphasize that where you need various perspectives,

00:42:44where today you would say, diversity in the room and in the chat and in the discussion is helpful.

00:42:50At this point, I would also say, we should work with several agents in different people

00:42:55You need to work on a topic that comes up in discussion, about the ecological

00:43:02balance and about how much energy is consumed and whether we should also perhaps

00:43:07switch models and not just run everything on the same manufacturer.

00:43:12I believe that in the cloud, the technical aspects that can also

00:43:17lead to this, but yes, then let’s maybe

00:43:23really tie a bow around this whole topic

00:43:27and very quickly to the team member, right?

00:43:31I believe, now once again for us as humans,

00:43:37the discussion we had with René,

00:43:40among other things about what the future of work looks like,

00:43:43how work actually looks, and also these fears,

00:43:46that are out there. I would now like to advocate,

00:43:49for all the people whose social skills are very good, who might not necessarily be

00:43:57the most tech-savvy, but I believe that if you can handle people well, you will

00:44:05also be able to manage these different personalities of AIs well in the future.

00:44:09And I believe that is a skill that remains important to us as humans,

00:44:13that this discussion we just had,

00:44:18as always in our episodes, where there isn't a right answer, but rather the right discussion

00:44:24maximally about looking at who we deploy, who do I need in my team as

00:44:32a member, as a human member, as an AI personality, or even as an AI advisory board

00:44:38in addition, to ensure that work is done well in the future and with good quality,

00:44:45these skills, I believe, are totally necessary, that one says one is empathic, one can deal with

00:44:51people, one can handle the opinions that are then formed, including these opinions

00:44:56critically question. I believe that the whole soft skill topic, which I've always somewhat

00:45:02been drifting along with, will interestingly perhaps become one of the most important

00:45:07skills in the age of AI. I think we can wrap up this episode with that.

00:45:17From that perspective, great to have you back.

00:45:20I hope that we will now return

00:45:21to a stronger regularity of our episodes.

00:45:24I think I have also received messages from some listeners

00:45:27along the lines of,

00:45:29Frank, is it continuing?

00:45:30Is Jens doing well?

00:45:31Yes, Jens, are you doing well?

00:45:32From that side?

00:45:33Thanks, Jens.

00:45:35It was nice to be with you.

00:45:36If you enjoyed it, spread the word.

00:45:38Leave us a star, subscribe to our channel

00:45:42and then I hope to see you soon.

00:45:44Jens?

00:45:45Ciao. See you soon. Ciao, ciao.

00:45:52Welcome to ThinkDifferent, ThinkAI,

00:45:56the podcast by Mark and Jens.

00:45:58Two technology-loving minds,

00:46:01who not only talk about artificial intelligence but live it.

00:46:05Here you get clear classifications, real practical insights

00:46:09and a fresh perspective on what is possible.

00:46:11Understandable, critical, and always with a wink.

00:46:15A.I. to ponder, to smile at

00:46:18and above all to discuss.