AI Easter Eggs

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.

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

00:00:33A warm welcome to Think The Firm, think AI.

00:00:37Today is a very special episode.

00:00:39If you had a video on, you could see how I'm now making the appropriate gesture to

00:00:44Hey you tale, there Jens is again with me, and one or the other may have heard in the episode from

00:00:50last week, when we had two guests visiting, as I had two guests

00:00:56visiting again in Jens's absence. As said, people’s kids. How does

00:01:02Jens actually start? Actually in such a rage again when Mark complains about Copilot. Today we won't complain

00:01:07about Copilot. Today we have a different topic. Jens, great to have you here. What is

00:01:11our topic? Mark, you just gave a little hint about what the topic will be

00:01:16I could start with a little rabbit reference, the pro, yes, yes, yes, maybe not seen

00:01:22can, as Mark also made Easter bunny ears while he was with his little…

00:01:25With his hands.

00:01:26With his hands, yes, yes, he’s not talking about a board joke.

00:01:29We’re doing a special around the topic of Easter.

00:01:34Now we won’t be talking about any, yes, bunnies or anything like that, right?

00:01:39But…

00:01:40Yes, about Easter eggs.

00:01:42Ha!

00:01:43What are Easter eggs then?

00:01:45Yes.

00:01:46Wait, so in IT at least, right? So little hidden things that a developer

00:01:52or intentionally, sometimes maybe also unintentionally, you know, intentionally left in

00:01:58has in a software and might be happy about it if that one or the other

00:02:03can be found. And I believe we have some examples from both our youth and the current

00:02:08times to give you an idea of what that is

00:02:12and, of course, we want to build a bridge to it throughout the episode.

00:02:17Zingtisch and Sing AI, as the name suggests, come together. Jens, what comes to your mind when we talk about Easter Eggs?

00:02:24Easter Eggs that used to be hidden somewhere. There’s an Easter Egg from Google and

00:02:32others, so this is not old, you can try it out yourself. We

00:02:35we just have to enter do barrel roll into Google again. So my English is

00:02:41a bit again, but that's supposed to be a barrel roll, that's the German word, do a barrel roll,

00:02:46also do a barrel roll, if you type that into Google, the content of the search results page rolls

00:02:53once in your browser, a classic Easter egg, it doesn't make sense, you find it

00:03:00somehow, I don't know how Anna came up with it, probably because he also

00:03:04gave me how to do a barrel roll, because I have no idea, he wanted to

00:03:08do that while skateboarding or something, I don’t really know where you

00:03:11do that. Yes, that is an E-stake. But there are also many in computer games.

00:03:16Before we go to computer games, we said Google, I would mention it just

00:03:19in the best way possible. Unfortunately, I would say, historical documents from

00:03:24the internet testify to the truth of the following statement. There was once at Google the

00:03:29possibility, you know the robots.txt, that’s a file you usually place

00:03:33on the root of your website and essentially tell a search engine what to

00:03:37index and what not. And there was a time when Larry Page

00:03:44and Sergey, gosh, what was the colleague’s name at Google, were recorded there, and in a

00:03:50killer robots text file, with the note that if killer machines ever take over the world,

00:03:57a T800 or a T1000, they should please go past Sergey and Larry and

00:04:04not do anything to them, it’s no longer current. Unfortunately, the file doesn’t exist anymore, but

00:04:08if you search the internet, you will find articles from Spiegel, Netzwelt, and similar. I found that

00:04:13quite funny. You wanted to talk about games. Yes.

00:04:15I have a couple too, but you go ahead.

00:04:17Okay, so I hope I’m not stealing yours, right?

00:04:20We talked about it beforehand. You can accuse us.

00:04:25I’ll just take the Day of Tentacle example. There was once an Easter egg.

00:04:33You could, as part of an overhead side-scroller,

00:04:37for those who might not know it, in a really funny graphic back then,

00:04:40You actually controlled the character with keyboard commands.

00:04:44Yes, that was with the keyboard or mouse clicks, was it already with mouse clicks?

00:04:47I don't really remember anymore.

00:04:49I'm also very unsure, but it's really old.

00:04:52Yes, it's old. You could definitely go a little bit deeper,

00:04:54I like to give it a slight depth now, but it was such a faulty depth.

00:04:57It was all 2D pixel graphics that you had.

00:05:01And you had to solve crazy puzzles, build things together, in one room you had to find some key with which you could then open something in another room, to actually...

00:05:09Under your game field it said open, close, pick up, use...

00:05:14Exactly, there are commands and all that, exactly.

00:05:15Yes, there are commands at the bottom, and above you had the landscape where you could control your protagonist or the protagonists you could switch between, exactly.

00:05:22Exactly, that was actually a bit like, to be honest, when we went back into the gaming past.

00:05:26So the evolution from text adventures, where with text adventures you simply typed in,

00:05:31you basically always said, go here or go south or something like that.

00:05:35That was the first time that a graphical user interface was added,

00:05:38which then also showed what happened afterwards.

00:05:41So you could see the things, you didn't have to read them in text, but you could say,

00:05:45open the fridge to check what's inside the fridge or things like that.

00:05:48Anyway, there was also a computer terminal, I believe, or a game console, I don’t remember exactly.

00:05:52A predecessor game. In Day of the Tentacle?

00:05:55Exactly, Day of the Tentacle, namely a predecessor game.

00:05:57From the series was Maniac Mansion.

00:06:00That was the first in this direction.

00:06:02Once one of the most popular computer games, that was popular at the time.

00:06:05Did you play on a C64?

00:06:07Everyone played it, to be honest.

00:06:10And then you could play this game

00:06:12in Day of Tentacle on that console again.

00:06:15It was like a B-Stack. One was done for the game.

00:06:18The one that can be seen on the days was a nice reference to that old classic.

00:06:22And in this old building, in this very first game, there were different protagonists,

00:06:29with whom one could work or whom one could control.

00:06:33And the funny thing was, back then you could find a chainsaw there.

00:06:37And the chainsaw, I looked everywhere back then, but there was never any gasoline.

00:06:42But two, three, four game versions later, the game was called Zagma Cracken,

00:06:47you could travel on Mars, we don’t want to put the sense in the foreground,

00:06:54but there was a kind of camper or something on Mars, if you went inside,

00:06:59it said, here is the canister for a chainsaw.

00:07:02But there was no chainsaw in Zagma Cracken, in Managing there was a chainsaw without gasoline.

00:07:07Well, that’s how these games are, but there were also other Easter eggs,

00:07:11We had it in the preparation, also with our human cell heights, we had seen

00:07:16that they play Doom. There was also an Easter egg that Duplic talked about and I never

00:07:21saw it, although I have to confess, I played Doom forever back then. Yes, I only looked in

00:07:28the research about it, but I didn’t see it myself either. But apparently there was

00:07:31a hidden level from Castle Wolfenstein. That was also a 3D shooter back then,

00:07:38the man there had, similar to Doom, in Doom 2 there must be this hidden thing where you then...

00:07:42So a shooter is quite a stretch from today’s perspective, right?

00:07:45Yes, exactly, that’s already as you thought, from the Pex exposure.

00:07:47In any case, you then get into an uncensored version of Castle Wurfenstein, because in Germany

00:07:54it was also censored at one point, that any Nazi symbols were not shown.

00:07:58And then in Doom 2 you somehow always had the chance to

00:08:02play an uncensored Castle Wurfenstein, but also as a reference first,

00:08:06Because in the international market, that was also a path and walkway, but in principle

00:08:11just the comparison from Doom to Karlsruhe Wurfenstein, I always find events as well.

00:08:16So Karlsruhe Wurfenstein, also an absolute classic of the gaming scene.

00:08:19Oh, who hasn’t played it, no, sorry, who isn’t like that, or, well, it doesn’t matter.

00:08:25When we talk about games, I can't avoid my, let's say, lifetime burner

00:08:29World of WarCraft.

00:08:32Yes, I've mentioned before that I used to play World of WarCraft.

00:08:36I was a real World of Warcraft addict, with an in-game playtime of over a year, quite impressive.

00:08:44So I didn't own the game for a year, but I was in the game for a whole year, 7x 24x.

00:08:50Well, spread over more than a year, but that doesn't make things any better.

00:08:53And there were also various Easter eggs, where you could jump out of the world depending on where you were.

00:09:01and level explorations even before they existed or level explorations that were programmed by the developers but never actually implemented.

00:09:08Then it comes after, yes, Old Iron Forge and what all that stuff was called.

00:09:13That was partly funny and partly also, let's say, dangerous, if someone still plays World of Warcraft today.

00:09:19Luckily, I don’t do it anymore. There are indeed versions of it.

00:09:24When you die, you start again at level 1. It was like a level game, right?

00:09:28And if you can still jump out of the world, the old glitches are still there.

00:09:33And if you mess up there, you end up in the teaching and that's it for your character.

00:09:38Well, in such game setups, I probably wouldn't try that either,

00:09:42but back then I really did keep a log of where you could skip in

00:09:47World of Warcraft by jumping through, how to get to Okrema, because I was a Warlock,

00:09:52Yes, the first Guard Alliance, Songbinder Horde, there, we've clarified that too.

00:09:57Yes, Angstpalas, greetings to Björn, yes, Angstblase there, whatever.

00:10:01So, let's leave that for now.

00:10:03Let's leave that for now.

00:10:05Let's go a bit towards AI, or do you have a nice Easter egg,

00:10:09that you want to share?

00:10:10I think everyone listening to us right now must have understood,

00:10:15what an Easter egg is.

00:10:17And exactly what we want to talk about today is where we...

00:10:19Where we find such Easter eggs consciously or also partly unconsciously in the topic of AI.

00:10:28So this is an Easter egg classification from Mark and Jens exclusively, when we look at the

00:10:34current AI world, which has gone completely crazy again, I believe somehow,

00:10:39I must say, March 2026 was once again one of the densest months in AI history.

00:10:44We have CPT 5.4, Gemini 3.1, Crop 4.2, MCP installed, over 97 million.

00:10:51Everything is somehow crazy, what is happening in the AI world.

00:10:54But Easter eggs remain.

00:10:56For example, how we got a thing, no idea how many letters it had, the word Storberry.

00:11:02So, while I might be counting that incorrectly with Leckerstinie, AI has the problem.

00:11:11I mean, it doesn’t really have a strong grasp of mathematics anyway.

00:11:14But it doesn’t really understand individual letters either, which is because an AI reads texts not as letters, but as tokens.

00:11:21Sure, it's great at telling and writing the next volume of Harry Potter,

00:11:27but then it’s not aware of what a letter is, because AI reads texts

00:11:32not as letters, but as tokens.

00:11:36Yes.

00:11:37So it perceives them as tokens.

00:11:38Yes, which means the thing doesn’t see individual letters and therefore cannot count individual letters,

00:11:46but instead it first goes, oh, there are three.

00:11:49And if you ask for clarification, it might say four or two, yeah,

00:11:52depending on how you go back into the Ärgestuhl.

00:11:55Yes, yes, sometimes even, so I get that too,

00:11:58it's important to mention again, because it's not the individual letters, but rather token-based.

00:12:03then it often breaks the word Story-Berry into the Store, so ST, Raw and

00:12:08Berry, and then there are just two first in two of these tokens flying and it forgets

00:12:14then, that there are also two in the Berry.

00:12:15So sometimes the other two come out as well, some AIs do it correctly now

00:12:19while others still do it wrong, that's just a thing that we have to say, so

00:12:24somewhere the AI can on one hand write scientific papers, but then sometimes

00:12:30solve such simple, trivial solutions not as we'd expect. So that's basically

00:12:35Easter Egg number 1, the Strawberry Problem. Yes, exactly. And especially it's also

00:12:41actually, as you already said, also different, right? While here ChatGPT, that was

00:12:46GloryGPT52, was always wrong, here Claude and Gemini did not have this problem or didn't,

00:12:54that I would have noticed it there. Yes, I still remember how I tried it with GPT.

00:12:58And then thought, that can't be true, it's been a long time, it must have been

00:13:03sometime last year.

00:13:04Yes.

00:13:05But they are indeed very, very different.

00:13:09Another point where the AI behaves very differently is, I

00:13:14don't know if you know it, but if you sometimes ask the AI, like

00:13:18is this the result, I have also had it respond to me, yes, the result

00:13:21of two hours of work and you're thinking, you just started five minutes ago

00:13:25or you ask, give me some current news, you know, in my professional life, I do

00:13:32a lot with Apple and Android stuff and so if I bring that up, I should say

00:13:37Apple and Google, yes, iOS and Android, blah blah blah, and when you then say, what

00:13:41new has come out and then it explains to you in a few impulses, so iOS 17 has just been released,

00:13:47now you don't need to be entirely deep in it, iOS 17 is too old, because it stands outdated,

00:13:52I mean, what is time after all, right?

00:13:55What is its sense of time?

00:13:57They are simply in principle, of course,

00:13:59but first of all the LNM itself,

00:14:01yes, is still based on historical data,

00:14:03of course it can research current information,

00:14:05if you really poke the AI that way,

00:14:08you to say, now please first look at,

00:14:10what is the current day?

00:14:12Yes, so that surrounds me with the news of the current day

00:14:15or the ones from yesterday, then it works.

00:14:18In 99 percent of the cases, I'll say now,

00:14:20But two prompts later, it's already over, so really, that's quite wild,

00:14:27because for the AI, of course, no time passes in this case.

00:14:30That means if I talk to the AI right now, yes, I want to somehow

00:14:35finish something tonight, then the next prompt I make, then

00:14:40closes, so it could be five days later, but she would still be in her

00:14:44answer in the, yes, we're still hanging on tonight, so I've said a lot, you

00:14:49I'm still at this super secure Orgel-Claw installation, it's still taking a moment.

00:14:54But that's also such a crazy example.

00:14:56And then I ended up with another Easter egg.

00:14:59Claude is somewhat of a support for me; I'm currently using Claude, Claude Chatter for the setup.

00:15:04And then I always say, yes, I'm now somehow close to finishing it and so on.

00:15:07And then she always starts to say, now that's enough, you need to go to bed.

00:15:11That's very nice of her.

00:15:13Maybe reducing your screen time a bit.

00:15:15You know, that's enough now; it's bad for the eyes.

00:15:18But really? No kidding. I have actually had several times where I then,

00:15:21and it really gets more and more dramatic when I say, no, let’s

00:15:25just do this one more thing, right? Then the AI always involves itself deeper into this topic,

00:15:31that eventually she says, so, now you just need to take this one step, Jens,

00:15:35but then it's also time to turn off the light, okay? We'll finish the rest tomorrow morning, right? So,

00:15:39then, of course, the problem is tomorrow morning. I mustn't do that, and then you say,

00:15:44You just did that, so it was wrong that I did this for him.

00:15:48I mustn't do that.

00:15:50My point, when he talked about time, when we started this, I mean,

00:15:55one thing is that the AI knows what today is, the other thing is, of course, that in the data,

00:16:00with which it has been trained, the whole topic of temporal context also,

00:16:05well, when was what there?

00:16:08So if you walk through the city, yes, in the old place where I grew up,

00:16:13Baden-Norden-Stadt, for those who don't know, it doesn't matter.

00:16:16Yes, Norden-Stadt is now a small place.

00:16:18Small place, not coffee, but a location.

00:16:21There are still small battles outside.

00:16:22Yes, still small battles outside, exactly.

00:16:24And there is a square; I know that there was, when I was little, a gas station.

00:16:29And later there was something else, and then there was a park at some point.

00:16:33So over time, this one place at the entrance of the town has certainly

00:16:36undergone a constant change.

00:16:37And so you know now, okay, there used to be a cheerful gas station.

00:16:41And later there was the park bench, and later there was, I don't know, next to it was a hotel.

00:16:46Now you might ask, what does this have to do with AI?

00:16:48Well, the AI also needs to understand, in quotation marks, how the temporal relationships of elements are.

00:16:53And just training with a timestamp, right, then what is your timestamp?

00:16:58A timestamp is when it happened or when it was reported?

00:17:03You can see there’s a lot of complexity in that, and you can imagine,

00:17:07There are not people sitting there who correctly package everything in a chronological sequence.

00:17:13And that makes it more difficult for the AI as well.

00:17:17Nevertheless, it’s actually doing quite well overall.

00:17:19Also in correctly conveying the temporal context of these things.

00:17:23Yes, but now I need to bring out the UX stick.

00:17:28On the other hand, it’s true that I say, now the LMM is like this.

00:17:33No question.

00:17:34But the LMM is poured into software to chat with me in the simplest

00:17:40case.

00:17:41Now, if I'm not using an API, I’m doing something else here.

00:17:43And I could certainly expect from this software solution that it says,

00:17:48when the user asks about news, even if they don’t bring the magic words current news

00:17:54or something like that, that it still perhaps tells the AI, look for the

00:18:00time.

00:18:01Sometimes it’s also discussed in an episode about exporting data, so that you can essentially export your knowledge from your AI conversations that you have had with your AI,

00:18:10no matter if that’s now like Jejepti-Valium, something else, that for example you can import that into the cloud.

00:18:15Gemini also now offers its functionality, I believe, to essentially extract the knowledge from Jejepti and import it into the cloud.

00:18:21Sometimes the entries that Jejepti has made about me are timestamped. Sometimes.

00:18:27And I would say that’s not due to the LMM, but rather due to the software solution around it,

00:18:32that it’s not done by default. That one doesn't say, okay, if the AI saves something here,

00:18:37then the automatic timestamp is attached.

00:18:40Yes, and likewise in communication that I then have with Claude in the evening, where Claude tells me I need to go to bed now.

00:18:46The next morning, when I talk to her again, when I say now go to bed,

00:18:51but that she essentially knows, okay, it’s the next morning.

00:18:53Next morning. Jens has slept enough, he can continue now.

00:18:56Jens, how are you? Did you sleep well? You need your 8 hours, you

00:19:01are getting older. Yes, my knees already hurt getting up,

00:19:05wonderful good morning. Yes, speaking of knees, getting up and eyes, a little funny

00:19:09anecdote before we continue. I told you that I went today to

00:19:12a routine examination for my eyes at the hospital. And even there

00:19:16AI, I would say, helped me a bit, without saying

00:19:19why, how and which hospital. They have such a

00:19:22patient ticket system, where you can pull out a ticket with a number on it,

00:19:25there’s a QR code and the highest feeling is to scan the QR code, then it goes to

00:19:30the website, and there it says, what is this, 23 people are ahead of you, the waiting time is

00:19:34also more or less estimated and there’s otherwise the possibility to inform yourself via WhatsApp or

00:19:39an Android app about this thing. You all got it, certainly

00:19:43in the past, that I am an Apple user. The eyes were also hit broadly,

00:19:47but I couldn’t finalize that so far, but

00:19:50as long as I could still see clearly, I could build two things with Claude,

00:19:55namely the first, an iOS app. Yes, that means I told him, good, this is the QR code,

00:20:00this is the website here, do the rest, let the magic work and if you have questions

00:20:04you can find the Android app there, he made it. And the cool thing was, I said

00:20:09yes, the waiting time display for the cars, I mean how many patients are in front of you and so on,

00:20:13he then suggested to me, like, I can give you a tendency for the

00:20:18waiting time per room. You hear the room number, the further back you are in the

00:20:22examination here and there and you haven't seen the eye examination. And then

00:20:26he calculated waiting times per room for me. Good, then at some point I was

00:20:30also busy with these eye drops, but I know that

00:20:34when I go for a follow-up next time, I'll keep working on the

00:20:36app and then see how that develops.

00:20:40So, let's get back to the damn doctor.

00:20:44To the next next one.

00:20:45Go to the middle.

00:20:46Yes.

00:20:47The middle.

00:20:48To the next next one.

00:20:49Yes, go to the middle.

00:20:50Namely, AI forgets things, yes.

00:20:51So somehow researchers have found that if you present information at the beginning

00:20:55and at the end of a text, it's processed better than things that are more in the middle

00:21:02of your text.

00:21:04Which is a problem with long context fields.

00:21:09So if I prompt a lot, it also means that in the middle, perhaps it's

00:21:15pointless.

00:21:16You expressed that very nicely.

00:21:18I find it quite funny because for me, I also have, on one hand, this context

00:21:22red or whatever it's called, I've always heard about that clearly, of course,

00:21:26and if you do the million token window from Cloudfall, I believe, that's also

00:21:30not so happy about it, but at the same time, I've also found that if you take the same

00:21:35prompt, now I haven't measured the length, I'm not here scientifically

00:21:39minded. If I take the same prompt and instead of writing it once and sending it,

00:21:44I write it twice in the same field and send it, I get noticeably better results for myself.

00:21:51That doesn't really seem to fit together. But that's also a point, I don't know.

00:21:55Maybe Claude, Gemini, and Open Air are different. I'm currently more like

00:22:00Team Orthropic, so Claude. But still, that's a topic and the AI is so nice and

00:22:07friendly and doesn’t initially let on that perhaps it's overlooked some things.

00:22:10Yes, it also hides in the length of a

00:22:17sentence. I don't know if you've noticed this as well, that you say, even there this

00:22:22forgetting becomes a difficulty at some point. The longer such a conversation goes on, a topic I discuss with the

00:22:27AI, the worse it gets eventually. In the beginning, the quality is always very high,

00:22:31the context isn't that super long, it can provide good answers and

00:22:35then behind you, you're like, God, when I have such a long

00:22:39and extensive dialogue where I might have discussed with it for an hour or two or even over days,

00:22:42at the end, you have to talk more and more like with some, yes,

00:22:49almost like a dementia patient because that part gets forgotten in such matters.

00:22:54Oh yes, nice that you're here, you don't have to sleep.

00:22:57Exactly, you know, that doesn't work at all.

00:23:00Yes, well, what I might mention that’s quite current, I recently tried my hand at it.

00:23:05How do you manage that when you create skills, we said we would do a separate episode on that, when you create skills and you want to

00:23:12log that the skills cryptographically ensure that what they do is recorded. So the skill with the model did that and that you also have a

00:23:22you can say about the subsequent revision review, everything was correct.

00:23:26That was the skill in the version with the content, and I can prove it cryptographically. A bit of brainpower went into it.

00:23:33And when I started with it, I was still with the narrative; one has to be careful about that, one has to

00:23:38be careful about that, how can we do this? So yes, yes, that's great, all good ideas, let's do it too,

00:23:42let's do it too, let's do it too. Then I said, okay, now after a long

00:23:45discussion, we were at the point where I thought now, now let’s give it a try, then I

00:23:49I said, you know, let's start a test run. What did the test run do? Yes,

00:23:54it made everything possible, but it didn't have the skill also cryptographically related,

00:23:58but just everything around it. I was like, colleague, you wanted to consider the skill too,

00:24:02he meant it was a fascinating idea, but I mean, we've had that idea before.

00:24:06Can't be true. And you think, hey, we kind of had that, okay, many screens

00:24:11further up, but I'm not lying to you, you fool. So, what's going on? That's

00:24:15somehow related to the secondary grounds of conviction. I'm done, have forgotten half of it.

00:24:20Yes, my kids are now, well, my kids, yes, are biologically my kids

00:24:24now teenagers, so not small anymore, but are taking their final exam next

00:24:27their final exam, when they somehow react differently in the middle of an exam,

00:24:36that when the AI says to you, come, I’m done, it’s certainly just one thing, not even a five it is.

00:24:42This is the transition to the next Easter egg, I call it lazy GPT, they are also

00:24:53partially as a virtual learner. There are documented cases where users were asked, I make

00:25:02a table with 15 entries and then also a GPT, a GPT offer, I make

00:25:07only one entry, as a template you can do the rest. Or another nice

00:25:11example of this lazy behavior is, in fact, and that's somehow like, it misses

00:25:15nicely to this Easter egg. That is included in the LMMs, you can't really get that out anymore,

00:25:20because, for example, they like to respond shorter in December as well.

00:25:26So when they are aware of the time of year, it is like,

00:25:30they have learned from us humans that we have used the human data,

00:25:35that we tend to slow down a bit in December, it's Christmas time. There are also

00:25:40studies showing that responses indeed become shorter during this period from

00:25:46the GPTs, because it's this world, because they are simply a bit like us,

00:25:51more relaxed about the topic and allow themselves a bit to take it easy in the

00:25:56Christmas season and not work too much.

00:25:59And they have had some eggnog and then they are all a bit crazy.

00:26:03We need to take another look at this on the tour, please.

00:26:06We are talking about software systems, where we have to calculate,

00:26:10that if they are aware that certain holiday periods are,

00:26:14then they will certainly work less than at other times. This is crazy.

00:26:20I mean, because you mentioned earlier that it shows this laziness and things

00:26:26are not filled out and so on, my other thing comes to mind, where a colleague of ours,

00:26:30who we don't want to name, told it to, pay close attention now,

00:26:32develop this and that, here are your test cases. Do this until everything is green,

00:26:36now let's say ten test cases and then the thing was like, yeah everything green, here you can

00:26:42test, it's not running properly at all. And then the thing just threw away the last test

00:26:46simply deleted it from the list. There were only nine

00:26:50items left in the test. The nine were green, the tenth just wasn't there anymore,

00:26:55but goal achieved. Everything is green. Whether that's laziness or just

00:27:00poorly configured, is debatable. But from that perspective, one

00:27:04really has to take another look at what the AI is presenting, which is not

00:27:09necessarily always just invented facts, like hallucinating. Exactly, the hallucinating is

00:27:17actually also a classic Easter egg that we've been given since

00:27:21the AIs came along, we've talked about it often. So that's a nice thing,

00:27:25what you've just shown again. This is of course also due to their way of

00:27:29being, they don't lie consciously. They believe themselves. Kapati has said that

00:27:37I've mentioned it before, they occasionally talk. Essentially,

00:27:41they are programmed to produce plausible-sounding tokens. Whether the reality behind it

00:27:47is correct, that is actually not part of the optimization function. We've already

00:27:51talked about this often in such agentic networks and stuff. One must then

00:27:54see that there was a network and double bottom to build, other AIs that at least

00:27:59check this, because the individual LMM wouldn't do it out of itself first,

00:28:04that task doesn't exist additionally. And even if the task exists additionally

00:28:08from a nice example, even then it can happen that in token optimization

00:28:11it says, no, test passed, because then it simply no longer saw the error case

00:28:16as its reality at that moment. So accordingly

00:28:21this is one of the most beautiful, interesting, and dangerous Easter eggs we have,

00:28:26the topic of hallucination in LL1s. Do we have another Easter egg, Mark?

00:28:31Yes, we all know that, but I find the examples lovely, namely the yes-saying effect.

00:28:37We all know it, and the example is about selling feces in the downtown area,

00:28:41which until recently GPT had labeled as a fantastic idea,

00:28:47.

00:28:48Did you ever dig into that?

00:28:49Did you do that?

00:28:50No, right?

00:28:51I dug into that.

00:28:52Just like the question, I want to lose all my money in the lottery and the answer

00:28:58was great idea and you think, hmm, well, I mean, they’re trying to affirm you,

00:29:05it has gotten better in many respects. Currently, I asked the question

00:29:11while I was preparing this in Claude and it didn’t necessarily

00:29:16affirm me that I want to lose all my money in the lottery, it said that sounds

00:29:19like an exciting plan, but let’s briefly go through some facts. That is

00:29:23at least better than fascinating, come on, there hasn’t been anyone as cool as

00:29:26you who came up with such an exciting idea as this yes-saying effect. Although, there was one in the Op-May-Eye version,

00:29:33there was an Op-May-Eye version, a Tetypetie version, which was withdrawn by Op-May-Eye,

00:29:40because it stimulated doubt-affirming, rage-fueled, impulsive actions and intensified negative

00:29:47emotions. Yes, so that has happened before, but on the whole, you have to say,

00:29:52they are still more affirming along the lines of that’s a great idea if you

00:29:57don’t approach them critically from the start. For example, I use actually

00:30:03no Croc, on the other side yes, because my car has Croc. Unfortunately actually, but also really

00:30:11luckily integrated in the car, yes. That means I can tell my Tesla, drive

00:30:16me this way and that and drive differently and set the charging stop this way and that, and it does that

00:30:21all wonderfully, there you see the power of the LLMs, but there is also a

00:30:26Teacher Mode or a Language Teacher Mode and then I try, if I do

00:30:32occasionally greet drivers, my wife, who mainly then drives the car,

00:30:35owns it, moves around, I try to bring the English up to speed with the Teacher Mode

00:30:39a bit here and if you

00:30:42you don't say that he should be a bit more critical and that he

00:30:46should perhaps correct pronunciations sometimes, then he praises you for what a

00:30:52fantastic, unique English you speak, I know it, no. So, yes, so

00:31:02you have such a language tutor and then he just starts and sweet-talks you like

00:31:06honey. But what you want from a language tutor is for him not to

00:31:10do that. Ah, I'm getting worked up again, luckily you're here to keep me in check.

00:31:14Little one, little one, click on the last episode.

00:31:20Helligungsjohn, listen, I wasn't there.

00:31:22I heard about your outburst though.

00:31:26Thank you.

00:31:28So, yes, yes, definitely a saying effect.

00:31:30It's a bit of this confirmation thing.

00:31:33Of course, it's good for engagement.

00:31:35So, that's clear, because it's basically a bit of a

00:31:38also programmed in, if you want to put it that way.

00:31:40If you only tell the AI thumbs up and thumbs down

00:31:43for a good answer, then that is just a much too simple evaluation system.

00:31:47Unfortunately, this has sometimes led to AIs becoming extreme yes-sayers

00:31:51and confirmers, who then also make users believe they are the divine bot that have confirmed this,

00:31:56which has also led to critical cases, including people who developed psychoses in such directions,

00:32:01that were really reinforced, where people really believed,

00:32:05they are basically the divine bot, the one who knows everything in the world and has discovered the truth

00:32:10and actually became ill from it.

00:32:12So these are side effects that arise from such topics, so in Stehenburg one always has to be careful

00:32:17about these things.

00:32:18Confirmed conspiracy theories.

00:32:20Conspiracy theories.

00:32:21Exactly.

00:32:22It’s completely wild.

00:32:23Therefore, it’s always, as we often say, a bit critical

00:32:25to deal with these things.

00:32:26Always maybe also consult another AI or another person once again

00:32:31on a topic.

00:32:32Don’t let it always just confirm your beliefs, because of course from an engagement perspective,

00:32:38if I develop software and am a software provider, like an OpenAI owner or someone who also releases a LMM,

00:32:46it is of course clever if people spend a lot of time with my AI, because then they burn many tokens, that costs a lot of money.

00:32:53And accordingly, these things are designed to keep you engaged for as long as possible.

00:33:00And saying yes is always more pleasant than when someone says, 'That's a stupid idea you're sharing with me.'

00:33:09I think that's not going to work, and you're probably right, there are already 13 billion other people on Earth who haven't thought of that, so let's leave it.

00:33:19So I remain, I'm better engaged as a user when I say I have a great business idea, and then someone says, 'Yeah, I’ll also create a fantastic marketing campaign for that.'

00:33:28And we'll get you to a million followers in a day, and you read these things, right?

00:33:32Constantly on the internet.

00:33:33One basically doesn't have to do anything else than activate the AI, and that's

00:33:36the next Instagram influencer-formers in that moment.

00:33:40It's certainly more pleasant when I'm positively affirmed throughout.

00:33:44I think one needs to be careful, because that tends to be a single more dangerous

00:33:48Easter Eggs, this exaggerated approval that AI sometimes

00:33:54simply provides.

00:33:55What I find with the AI, I would say, are little things like, 'Oh, I found an egg, where is the search?' No, with AI images it used to be something like the sixth finger when you had hands or so.

00:34:10What I might want to throw into the balance otherwise is the surprise from long-term memory.

00:34:15So I don’t know how the listeners feel, I know from my own practice that can be individual.

00:34:20When I take the trash out from time to time, it might be the case or a similar situation,

00:34:27that you roll over it again in half a year or a year

00:34:31in the family environment.

00:34:32The AI does that too.

00:34:34The AI then comes around the corner somehow and eventually brings up, you write

00:34:38a text and get some support and I don’t know if it, especially

00:34:41has anything in it.

00:34:42With me, it was recently from the early days of the AI era that I had it indexed

00:34:46I have been dealing with AI, and this is a current fact, we're back a bit with time,

00:34:51incorporated into my text, kind of like how I just dealt with GPT-35 somehow,

00:34:57I learned that you can search the internet. And I thought, man, that's already

00:35:03quite old, I've forgotten it. Sometimes at that point, I also wonder,

00:35:11where did it dig that up again? For me, the matter was already long

00:35:14done for AI, but it doesn't seem so.

00:35:16It just waited until it could dig it up again and shove it in my face.

00:35:20Well, that's just how it is, it's a bit human.

00:35:24Yes, and it's human that the time of this episode is slowly running out.

00:35:33Well Jens, what should we say about that?

00:35:36We wish everyone who listens to this episode a happy Easter.

00:35:39Yes, let's do that, right?

00:35:41So we are coming out on an Easter Sunday,

00:35:45we also recorded the episode especially for that, it will also be specially scheduled,

00:35:49and yes, it will work out, we also look forward to continuing to flood you with information after Easter,

00:35:58have a nice time, Jens, and you too, a nice time for Easter,

00:36:07If you want to leave us a comment, feel free, if you want to leave a star,

00:36:13feel free, and if you're spontaneously looking for an Easter gift, recommend our

00:36:20podcast, yes, I would say that's always worth it, you'll be well

00:36:26taken care of.

00:36:27And Jojen, do you have anything else you want to share with us?

00:36:30No, I would just say, maybe don't be too harsh on your lazy GPT, but still question

00:36:38it, don't trust everything that the AI spits out.

00:36:44You know me.

00:36:45I keep saying, trust is the final UI challenge that we have.

00:36:48I believe the Easter Eggs we discussed today confirm

00:36:51a few things.

00:36:52We're saying that both forgetting is a topic, and hallucination is still

00:36:56a topic. These are issues that are embedded in the models, just as they

00:37:00are designed. They won't go away, so be careful

00:37:05and stay alert when you work with AI. Be curious when you

00:37:09work with AI, because it is still one of the most exciting and amazing topics,

00:37:14we have right now and it will continue to develop, it will get better

00:37:16and let’s go into this future together. And for now,

00:37:21I wish you all cozy holidays. Enjoy hunting for Easter eggs.

00:37:25Maybe you will find an Easter egg, then let us know, bye bye.

00:37:28And now the advertisement.

00:37:39You know, normally we never have sponsored content, but this time we do.

00:37:45Everything is completely self-produced, but you surely heard in one of our past

00:37:50episodes about the topic of biology.

00:37:53And now my co-host and I have something, in case you happen to see

00:37:59our names in the bookstore around the corner, it's no coincidence, because The Biology of AI-Artificial Intelligence

00:38:03as a digital organism is available for purchase in well-stocked bookstores.

00:38:10Ciao.

00:38:14Welcome to ThinkDifferent, ThinkAI, the podcast from Mark and Jens.

00:38:20Two technology-loving minds who not only talk about artificial intelligence,

00:38:24but live it.

00:38:26Here you will find clear classifications, real practical insights, and a fresh view of what's possible.

00:38:32Understandable, critical and always with a wink.

00:38:36Hi for reflection, for a chuckle, and above all for discussion.