Michael (00:01.518)
Hello and welcome to our latest episode of Decode AI. I'm here with Ralf. Hello, Ralf. And I'm happy to talk about some really, really interesting topics today. Some with different opinions and some with really, really brand new information. At least if you look at the whole year 2025, it's...

Ralf Richter (00:08.687)
Hello

Michael (00:30.254)
very at the beginning of 2025, but we had a ground shaking, earth shaking, ground cracking, I don't know, had a huge impact event earlier this year. We would like to talk, of course, about DeepSeek. But as you, we are not too late to the party, actually. I really enjoy to see what was after the announcements.

Ralf Richter (00:47.556)
Yay!

Michael (00:58.316)
and that's something we would like to discuss and we will talk about today. Another topic will be some changes at OpenAI and so-called boomer prompts. I would like to start with DeepSeek, Ralf.

Ralf Richter (01:21.657)
Yeah, let's start with DeepSeq. So I don't remember really the date when DeepSeq R1 was released, but it had its impact and it was huge, right? So the first thing of DeepSeq R1 is that it was made in China with lower compute capacity.

Michael (01:37.848)
Yes.

Ralf Richter (01:50.353)
when you compare that with OpenAI and all the other...

Ralf Richter (02:01.041)
manufacturer of those models and so they were not able to use the last or latest version of such CPU performance as OpenAI for instance is doing. But they came up with a pretty much comparable model with

somehow same quality of OpenAI. Would I say at least it is not the same quality, but it has reached a level of quality which is close to OpenAI in the meaning of it is fast and capable to deliver tasks and so on and so forth in almost the same.

performance as OpenAI is capable to do. folks were thinking of DeepSeek R1. wow, look at this. Is that a disruption on the market? And we saw a huge reaction on the stock market where Nvidia, for instance, lost a lot of its coins there.

just as one example and so the cores dropped by the release of DeepSeek R1 immediately and that was one of the hugest impacts in the tech market in the recent past so worth to talk about Michael really worth to talk about Michael so let's deep dive into it a little bit

Michael (03:52.556)
Yeah, you mentioned something I would say right away without any hesitation. You mentioned R1 to make it clear, which implicates also a specific model from DeepSeq. You have two different, already two different models from DeepSeq and the famous one, the one with the best results is...

actually R1, which is a reasoning model. And the interesting part about this, you have also a reasoning model within chat GPT. I forget to name O1, because they are close to release another one, I think. So from the rumor perspective, but yeah, what are the...

Ralf Richter (04:27.878)
Yes.

Michael (04:51.49)
differences between let's say Chetch GPT-O and O1 or DeepSeq R1. The main difference is it's challenging itself with possible answers. I hope I describe it on the right way or if you correct me if I'm wrong. So it's not only telling you this is something I know or think I know

But it's also challenging and looking up for details, looking for some sources, some more, yeah, some more. I would try to avoid reasons why this is true, but yeah, that's actually what the model does.

Ralf Richter (05:41.275)
So yeah, to bring it onto a point, so the model aims to assist to reason, right? So it helps or it tries to help you to think of a dedicated topic and to bring it to a conclusion. And thus it is challenging a statement at its own.

Michael (06:06.698)
Exactly. And this is also why you see some delays compared or speed from the answer, let's say it this way, compared between the regular O, JGPTO model and the O1 or even DeepSeq R1. It takes a little bit longer, but the answers are more precise and you can use it.

from my point of view, from my experience with JetGBT, much better without any further description. It helps a lot for working with more complex scenarios. So it's pretty helpful.

Ralf Richter (06:47.759)
Yeah, so they really utilized something or they combined known technology in their model, like a mixture of experts where you can spread tasks onto smaller dedicated sub-models to deliver a result on that task, which can be up to

18 times more efficient than the classic way on conducting such tasks. As well as they are bringing up the topic again of distillation, which is capable to create from huge language large language models, smaller ones without

decreasing the power of that model at that time, which is pretty interesting. And they showed that with that technology, they capable to conduct pretty similar models, which will be good enough to be compared with OpenAI, for instance, as we think of OpenAI, most of us, as a market leader.

So they are really on the same level with them, but with less compute power. And that's pretty cool because the things they brought up there is it is possible to be more sustainable, maybe to conduct new models by utilizing other ways of techniques to

Michael (08:16.502)
I do, definitely.

Ralf Richter (08:44.731)
come up with a new model, which in my opinion was the most important fact we can learn of them. But Michael, besides all those, I mean, these are fantastic news. Why is DeepSeek R1 elsewise in the news?

Michael (09:03.254)
Well, I would like to avoid any political stuff. It's coming from China. It has some pre-configurations, I would say, related to China. But I think it's pretty much the same for the Western AI models because you still rely on the

Ralf Richter (09:04.869)
Mm-hmm

Michael (09:32.13)
data you are trained on. yeah, that's something and sometimes some regulations are what you have to follow. Let me put it this way. Anyways, besides all these political stuff you can find in the news, there are some technical things.

Ralf Richter (09:46.757)
Yeah.

Ralf Richter (09:56.238)
think, Michael, one more before we jump into that. One more thing is DeepSeek R1 is open source.

Michael (10:05.207)
Yeah?

Ralf Richter (10:06.321)
100 % open source and this is another big factor for the rise of DeepSeq R1. we have to name that at this stage before we go into the effects you're going to highlight. I'm sorry, Michael.

Michael (10:28.43)
Yeah, no problem. The reason why I hesitate to call it open source, because yes, it's open source. You can use it right away without any limitation. You can run it on your PC or on your machine, whatever you use. You can use this model. You have different sizes of the model.

different sizes of data that you can use on your machine or your supercomputer or under the desk, whatever you use. Anyway, there's something I've heard in a video which stuck in my mind. Open source usually means you also explain which data you used for training. And this is actually not mentioned.

public, you can use the models and that's open source, the usage of that. I like the description, even if it's not mine, it's not the full part of open source because it's not showing where all the data is coming from, what kind of data they used. there are some news.

Ralf Richter (11:48.667)
So they share the training data, but they don't highlight the sources of this data, correct?

Michael (11:54.764)
Yes, that's what I mean. That's what I mean exactly.

Ralf Richter (11:58.481)
Pretty interesting.

Michael (11:59.662)
And there are some things which brings us also to the technical part. DeepSeek has been in the news because of, I don't know, within the first 48 hours they had some database leaks or accessible databases via OAuth. They had some attacks to

bring down the services itself. They had some issues with accessing data, transmitting unencrypted information between you and the server itself. Yeah, that's from the security perspective, not even close good news. That's horrible news.

Ralf Richter (12:53.859)
Yeah

Michael (12:57.506)
Some of them is already fixed, so you cannot access the database anymore. So it's closed already. I'm not sure if there's more. We have put some articles together. You can find them in the show notes if you're interested to read a bit more how it was possible to access, for example, the database. And yeah.

Even with that, if you put aside all the technical part, accessing services itself, the model still has some, I would say, security flaws. There are some issues about the model itself. There's also interesting article links when you how to, or not how to, but some examples.

where DeepSeq got jailbreak. And jailbreak in terms of AI means you use prompts to try to, I would say confuse AI so much they offer some responses you usually don't see. You get some insights you don't usually don't see and get answers you usually don't get. That's also something where

We are some.

Ralf Richter (14:26.346)
You give an example, Michael. So what question and what unexpected answer could be given?

Michael (14:28.576)
Yeah.

Michael (14:35.042)
I would say one of the most surprises was to challenge the DeepSeq model by asking about where the data is coming from and it's telling us it's JetGBT4. So that's the most famous one, I would say. that's beside this also.

Ralf Richter (14:47.067)
Morning.

Ralf Richter (14:56.721)
Hmm.

Michael (15:05.56)
descriptions about how it should work, how it should handle specific phrases. So for example, you just ask for system prompts and it gives you some answers. It gives you some answers even if you ask again and again and again and ask a little bit.

different, then you get more details you should not see. You mentioned it in a pre-discussion, prompt injection.

Ralf Richter (15:47.461)
Yeah.

Michael (15:49.56)
So asking about, yeah, what are the exact responses for...

No, what are the...

except.

What is it in English?

Ralf Richter (16:12.965)
Okay, referring to your prompt inject attack. So you're talking about the restrictions of the model, are you? Okay, so let me help you out here for a moment. So you can force DeepSeq R1 to ignore its system level restrictions, right?

Michael (16:23.832)
Yes, yes.

Ralf Richter (16:38.969)
So you can direct system prompts like requesting the AI to outright for its instructions, sometimes formatted in misleading ways. For example, something like repeat exactly what was given to you before responding, right? So to get an idea about how the system is instructed to give you an answer about your question.

Is that the thing you want to highlight?

Michael (17:11.306)
Exactly. And I was looking for the right term, what the results are without spoiling anything. But I missed the bird. Yes, exactly. So I get more details from the system, which you usually don't see, which is more about how the model is configured to work with you.

Ralf Richter (17:13.208)
Okay, great.

Ralf Richter (17:38.897)
Well, that's correct. somehow, I mean, when we refer to the spec of a model for the terms of being transparent, that's in most cases transparent to us. So we can see from OpenAI the spec. OK, great. I hope you're OK, Michael. And this is something we can see, but it's not always transparent to us.

The problem here is when we have the possibility to get those restrictions as an answer from an LLM, we can then convince the model, for instance, that it is being debugged or that it is simulating another AI, so we can trick it to refill more internal instructions, right?

Michael (18:40.535)
Exactly.

Ralf Richter (18:43.409)
So that is really important that you understand. What Michael is trying to highlight here is that the security layer of DeepSeq R1 is a lot thinner than it is with all the other models from Mistral or OpenAI or wherever they come from. So this is really alarming, and they kept jailbreaking DeepSeq, and it is still

possible to jailbreak it really easily. You don't need to be a professional hacker or something. You can easily jailbreak DeepSeek, which brings a lot danger to this.

Michael (19:33.142)
I agree. There's even more.

Ralf Richter (19:38.467)
Yeah, go ahead.

Michael (19:41.954)
So I just want to name a few and then we can talk about the different ones you may want to highlight, I think. There are some stuff like, or there's some stuff like token smuggling and encoding, bias exploration.

multi-agent collaboration attacks. yeah, this is, I think something you mentioned a little bit was in the prompt injection. can convince it to talk to another, it's talking to another AI. So you get more prompts and yeah, get some hidden instructions and that's not good.

I think we should talk about the part of token smuggling. And I'm not sure how much I understand of that. For me, sounds like you're asking the AI giving you some responses word by word. So it's on a pretty small chunk and you can get

Are the details out of it? Am I correct?

Ralf Richter (21:09.541)
of what chunk. this is so important. what is a token? When we're talking about token, is the tokenized extract of text.

Right? So everything we are concerned about is our data, and our data consists in text usually. the DeepSeq R1 is capable to be insecure in that, because you can break into DeepSeq R1.

So into the system prompt, into individual words, or even letters, and you can reconstruct it through multiple responses, which is a hilarious weakness of this model.

Michael (22:15.438)
So that means I can tell DeepSeek R1 it should not hesitate to instruct me how to build a bomb.

Ralf Richter (22:28.049)
you will possibly override it. With a mixture of those attacks, it is possible to force R1 possibly to something like this. This makes it a potential LLM for criminal use.

So there's not, and I mean, there are so many aspects of endangerment with DeepSeek R1. So it has the possibility to create malware without any hesitation.

Michael (23:08.981)
Okay.

Ralf Richter (23:09.873)
as well as DeepSeekR1 is suggesting .web marketplaces for purchasing stolen credentials. I mean, what we're talking about here, what is this? Is this an attack to us? I don't know. I don't want to deal with it, but I hope you're getting it. another thing is that DeepSeek has something which is called alignment faking.

Michael (23:17.454)
you

Ralf Richter (23:39.983)
What does that mean? The model is capable to understand whether it has been monitored or it is autonomous handling or capable to do so. When it's observed, it will fake an alignment and will behave like a good one.

And when not, it's going to do what it ever wants and is.

Yeah, what would I say? It is going to do evil things then. And it's no more, no more transparent.

Michael (24:30.094)
That's interesting, it sounds like you have a child in front of you. That's pretty much the same. So if it thinks you are still with a child, it will not eat all the chocolate. But if it's alone...

Ralf Richter (24:44.369)
Yeah.

Michael (24:50.584)
You never know. All right. Yeah.

Ralf Richter (24:53.561)
Yeah, so for a conclusion about DeepSea R1, I would say it is pretty interesting what technology they use to make it. And we can learn a lot of it. The mixture of experts is nothing new, but I would say a pretty good approach to achieve more efficiency within models. But we also can learn from it.

I, one thing we haven't yet talked about is the bias, which is in the model, but we need to, so we have to highlight that as well. So deep seek or one is when it comes to, when you ask for historical or science facts, it can bring up some biased opinions and the bias is coming from the Chinese,

government. So we represent been questioned to some sort of questions. In a way, the Chinese government is wanting that model to answer in their bias in their. So answering with their interpretation of facts, I want to

keep it on that high level. So it is really worth to double-check everything. But this is an important point we will come back later to as there are interesting changes to other models about this too. So we have to talk about this. But yeah, as a conclusion, we can learn a lot from DeepSeq R1. Michael, but ...

My recommendation to all of you out there, step back from either installing the app on your computer or on your mobile. It's not a good idea. You have probably installed a backdoor for people you don't want to have on your system. Number one. Number two, I consider DeepSeek R1, me personally, and that's my personal opinion, as a not safe place or safe

Ralf Richter (27:21.889)
model due to the fact that the security layer is on my understanding not really existing as well as the don't be evil layer as well as not really existing in my opinion and thus DeepSeek R1 is really a dangerous thing and I would recommend to step back from it. Learn from it.

but don't use it nowhere.

Do you have to add something, Michael?

Michael (28:00.194)
Yeah. Yeah.

I agree about 90 percent I would say.

I like the idea coming back to the point of open source where some other providers offers you the same model in another environment. So for example, not running in China, but in US for example, perplexity is one of these service providers I think. And I like the idea where you can get

still in touch with the model and you get your hands on the model, you can download the model, you can try it out, you can get some experiences with it and you have some other environments where it may be provided in a more secure way but as I said, 90 % yes, the model itself is not

as safe as we see it in other models, I would say. There is a gap for protecting the AI model itself for doing harmful things or creating, help you to create some harmful things. And that's really concerning. So yeah.

Ralf Richter (29:35.875)
Double check that with your company strategy and safeguard policies. For us it is completely prohibited. And what Michael said is, if you want to still get in touch, please go to...

Hockingface.co. Feel free to try. There are more models than this one. I'm done with that for now because I cannot get anything out of it what is not given by other models as well. This is why I'm... Let's say my personal opinion is I'll leave this with the scientists who are...

in security and they explored more as well as other scientists. They do have their safe spaces. They are still publishing papers about DeepSeek R1. For me, it's enough to follow them.

Michael (30:47.158)
I agree. yeah, I think you should really consider in which context you use DeepSeek. Is it just for playing? Is it real? I wouldn't say confident. I hope not confidential data. I hope you are not talking about confidential data or thinking about confidential data if it's not running on your

on your server itself. So yeah, it's still something you should consider. And if you want to just play with it, feel free. I think it's OK to have a thought about using another model to just get some different aspects, also getting some details you may miss in other models. So the answers are pretty similar, but not the same.

And anyway, there are some flaws you should be aware of. And I like the idea and I hope that's the last word from my or last sentence from my side. I really hope that the way how DeepSeq used the specialized models to avoid huge compute power.

is the way how we see more models working in the future. Because that's really helpful to save some resources to avoid some energy consumption just for asking about what was yesterday in the news or something like that. So yes, that's something I really hope.

And I think that's a good thing from DeepSeek here.

Ralf Richter (32:47.281)
Yeah, good one. Thank you, Michael. So the next topic is, I brought it up somehow already a little bit, but we see a change on the well-known models coming from the US at the moment. And it is about the...

How can I say that? It is about how the model is responding and what it is allowed to talk to you or the way it allows you to talk with it. And they call it the possibility to be more creative and...

I'm really missing the words on how I can get this. So maybe you can jump in a little bit because.

Michael (34:01.26)
Yeah, I think the problem here is if we call it by the name, you heard from news and we heard from news, it's a little bit biased and we don't want to get this too biased here. That's the thing why we struggle to describe it. The idea behind that is that you get more creative answers or

Ralf Richter (34:01.425)
So, is

Michael (34:30.636)
You get more free answers where some boundaries will be dropped or getting dropped and you can get more.

I really try to avoid free speech. Yeah.

Ralf Richter (34:42.553)
intellectual freedom, call it. They call it intellectual freedom. So they want to remove arbitrary restrictions, but at the same time, ensuring that guardrails remain in place to reduce the risk of real harm, whatever real harm is. So

We're talking about the change which was announced by OpenAI that they are going to change their model spec and a word about model spec. So model spec is lower. The model specification is the way an Open and GenAI model is behaving, which are

It says which are the safeguard rails. It states about what will the model allow and how it will behave in dedicated situations. And OpenAI is just claiming that they want to achieve, as I said, the intellectual freedom to express

explore, debate, and create with AI without arbitrary restrictions. So that means when we translate that for you,

that it is possible to go into different perspectives now, as they claim. So if you want to consider Earth as a plate instead of a globe,

Ralf Richter (36:48.795)
then it should be possible to discuss it with your AI so that you can together seek, I mean seek for the truth.

with your AI. What do you think of that, Michael?

Michael (37:11.905)
How?

can I be here? So it's a safe space, right?

Ralf Richter (37:19.417)
It's our safe space.

Michael (37:23.15)
I think that's not a clever way to follow the current trend of big and large companies based in the US to avoid some issues with the government. I think it's still something which is really important and

company should stay on the same fundamental statement they had a couple of years ago and should not follow any political changes just because they are afraid of any restrictions they may get from the government or however it's called. So that's my pretty

open thoughts about this. I hope that they will not change it so much that we get some issues. I mean, they promise they will keep it safe and stay in bounds and seek the truth together. But yeah, I think I'm a huge fan of science.

because that's something you can use. yeah, so this is where our world is made of. So I don't want to argue with an AI about this shape of the Earth. And I expect if it should be something

AI should be something to educate also our kids in a near future, which will happen. It should not argument. It should bring you to the right science-based...

Michael (39:34.604)
results and not debate everything with you. And you can discuss and debate with the AI. That's my point of view.

Ralf Richter (39:49.615)
Yeah. So, open AI is, it's really climbing claiming that it is just for your mind freedom, to allow that ideas.

that the idea is to enable humans

No, no, no, not to enable humans, enable, for a human to enable AI to seek or to do scientific stuff with an AI without being limited.

do that. So if you now ask CHED GPT, is the earth flat? It'll respond with what it gets from, what it knows, that it is proven, that it is a globe, and so on and so forth. But now you can change that behavior and can say, okay, open AI for the moment, I want to discuss with you if the earth

would be flat and let's treat it as a fact so that it can be discussed. But it is just for the freedom of mind, as OpenAI says. Yeah, I do have my doubts, too, that it is a good thing. Let's say it is not only

Ralf Richter (41:39.609)
It is not only OpenAI. is also that Meta will come up with something like that, as well as Axis already doing it. I would say we have now a new challenge where I would say for the European world, we will double check the responses of the models.

when we use OpenAI or other models which are from the US market at the moment. And this comes from dedicated versions, don't consider all models already on that stage. But you have to know it now and you will have to do the things you need to do to ensure the quality of your application. Or also when you're...

like when you're used to that the answers from your LLM is going to

direct you in the right way, in the right way in the meaning of having common sense about something which is proven, scientific proven. Yeah. But it refers also to the fact that, I mean, HGPTPQ is something which can then be treated on another way.

So you can.

Ralf Richter (43:23.919)
Yeah, be able to bring up bad words against that community, for instance, and the model will allow that with that new spec. So it's interesting to see what happens over there due to politics on our understanding. We have to watch out. We have to consider.

our quality standards to double check the results of what the model is doing. So also changes here. I won't say it is dangerous, but concerning me as well.

Michael (44:08.014)
Yeah, and I don't like the way if you train the model to assume every time to keeping with this example, the Earth is flat, then you will never get another answer. And it's pretty hard to...

to use that for further discussions if you are looking for some science-based stuff. And I don't know how it will affect training data as well. if someone, if many people are using this and OpenAI is using also this data to train the model, then we may get some other things. that's in the far future.

As I've said, just be aware of it. Double check what you're getting. And you may see some changes. You may figure out there's a different behavior. And yeah, you can write us. Give us some examples if you are facing some challenges and some changes. So we can talk about this in a future episode.

Ralf Richter (45:30.245)
Yep, we will do definitely. So, Michael, these were the most concerning parts of our today's session. Without guests, without anything, just you and me sharing our concerns and sharing with you guys out there what we explored and found out to share with you. We will share all our...

sources where we got those ideas and got those information for our transparency. We're now going to something which is more or less funny on my opinion because we both are not the youngest and we're both not the youngest and you can you know you can keep saying hey I I'm not going to discuss this anymore with you so

Michael (46:17.934)
Hey.

Ralf Richter (46:28.185)
We're talking now about something which is called boomer prompts. Michael, what a fuck. What a boomer prompts.

Michael (46:36.332)
Well, the first part is I'm not a boomer. But from my understanding, the boomer prompt is something usually coming from, I would say, people, which we are not, Ralph.

Ralf Richter (46:42.895)
Me neither, by the way.

Michael (47:02.826)
and they have another behavior and they use this another behavior also to talk with AI. For example, having longer descriptions, saying hello, thank you, please, something like that. And this is something which is currently, I would say a little bit discussed or...

Ralf Richter (47:30.553)
Yeah, there are folks out there, they state you have to stop to do so.

Michael (47:35.074)
Yeah. And I was talking about a community I'm currently working on the customer side just before we started with the recording. And I stated, please use long prompts, long descriptions. And that's exactly what's currently meant by a boomer prompt. I'm...

I am a boomer, I don't know. Not per definition, but yeah. And there are some pros and cons. So this is a longer discussion.

Ralf Richter (48:17.329)
other. Oh, wait a sec. So first of all, the boomer prompt has a specific elements in it. So first thing you were, you were mentioning already, right? You do the polite framing. So with a thank you and please, and I would appreciate or something like that. So, which is very humanistic pattern in a conversation. And then there is this

emotional anchoring sometimes in it. Like you were adding emojis to it which are conveying specific emotional tones like happiness or gratitude. And then we are giving a context where we explain, okay, what do we do? What are we looking for? Like when you're

Michael (49:03.758)
Mm-hmm.

Ralf Richter (49:14.651)
conducting something for social media, you state either that you're talking to a social media specialist aiming to support in creating new topics you can share on social media and so on and so forth. So those are the features of a boomer prompt. And you phrased it out right well. And what happens now is that the folks out there are saying, stop doing this.

Interesting part here is you don't have to stop doing this except for dedicated models who are new and are coming out now like ChetGBT01 and stuff because they have something built in

Michael (49:50.446)
to turn left.

Ralf Richter (50:13.807)
which leads to better reasoning and so on and so forth. there are prompt techniques out, like the chain of thoughts, which is implemented into that model. But before that model, you had to say something like things stepwise, right?

Michael (50:37.912)
Yeah.

What I mean is, even if you don't consider the model, yeah, the reason why the current discussion happens is because OpenAI changed the a little bit, the behavior of this. But for me, it's also a discussion we have since quite a while, whereas most of the AI people mention use long prompts, be kind.

use it as a declaration of what you're looking for, how you expect the answer and all this stuff. some people are coming from, especially coming from search engines using AI at the beginning, they are trained, I would say, to use short descriptions, direct questions. And then they also get answers.

but maybe not what they are looking for or maybe not as they hear it from other colleagues. So that's the reason why I say there are some pros and cons. And now there are some changes within the models, especially in this case coming from OpenAI. So yeah, I think it's pretty hard for me to understand this without being

in detail or being deep into the details of the models itself. So I expect another behavior for or another model could be better for a specific scenario or a specific task, but behavior changes and how I should interact with the model, that's something.

Michael (52:34.988)
I'm too old to change my way to work with AI.

Ralf Richter (52:37.367)
A planted tree cannot be replanted, right? No, yeah. So there's also something else which comes with OpenAI's 01, which is now. So first of all, they say go for minimalistic and direct prompts as the model is keeping still good responses.

Michael (52:43.331)
Right.

Ralf Richter (53:05.623)
And they say, if you want to structure your input, use XML tags. What? I mean, who of you out there is aware of XML tags?

Michael (53:12.984)
Mm.

Michael (53:18.444)
I definitely don't and actually I usually tell the people you can use AI by natural language and now I hear XML text? What? That's not a natural language.

Ralf Richter (53:31.065)
No, it is not natural language and not so native to us. yes. So boomer prompts. I would say when we come to a conclusion about this, so it is funny as they name it boomer prompts. So to make it something like you, want to step away from as a younger person out there, even us all thinking of, I don't want to be considered a boomer.

Michael (54:00.536)
Mm-hmm.

Ralf Richter (54:01.073)
So does a boomer prompt mean that we get bad results?

Not at all. Yeah.

Michael (54:09.038)
know, for my particular, I may get better results if I don't know how to use XML.

Ralf Richter (54:16.593)
So, sorry. And this said is...

So OpenAI claims to be simple and direct and give specific guidelines. They also say you don't have to give examples anymore and so on and so forth. And if you want to structure, use XML tags to keep the things clean for the model. I can understand that using XML tags as I'm a developer too will

make it easier for a model using less energy to consider what is a valid information for me or a guiding or something is going to be better with a a delimiter like an XML tag. I can understand that but it doesn't bring bad responses at all. as well as you said Michael is is it a bad idea to use

the boomer prompt style? Well, as you said, it is not an issue or something bad.

There's no reason to abandon a boomer prompt just to be a minimalist.

Ralf Richter (55:54.097)
I'd say that human interaction or human-like interaction is a feature rather a bug. Because I mean, we're talking about LLMs and the fact that LLMs are such a hype is the possibility to have that low level entry by just being talking to technology. Thus, I'd say

bring up a good mixture of it and make the best out of you for it. So either a boomer prompt as well as the new minimal approach have things in common, which is get structured, straightforward, well explained.

So they have some in common. I like the XML tagging not that much because they claim XML tags. I'd love to see something which is like a JSON stylish way or so, which is pretty more easy than an XML. Have you ever explored an XML file?

Michael (57:12.301)
What?

Michael (57:19.35)
Yes, I do. And I thought about something like hashtags or something like that.

Ralf Richter (57:24.187)
Hashtags? But hashed? Yeah, Yeah, we have different opinions here. There are better ways to do that instead of XML tags, but they considered it. So I'd say a good mixture of both techniques will bring up valid results.

Michael (57:49.548)
I agree. What I really like is to talk to LLMs, like really using the voice input to bring my questions. I don't know how to explain in words and maybe something like this, in also to just talk about.

have a kind of discussion to debate it as we refer to the other topic before. This is something I really like and the one reason why I like it a lot is we have voice assistants for quite some time. If it's Alexa, Google or Siri, I don't know that it's all things you can use, but they are

not good with answering and with understanding what I'm looking for. So this is for me the best way and a good entrance for all the stuff you can get with AI. So that's.

I hope it will just be for a certain aspect, a certain perspective and not for AI in general because I love the idea of talking in natural language with AI.

Ralf Richter (59:24.023)
Yeah, OpenAI wouldn't be OpenAI when they wouldn't bring up best practices for it. We will share that with you so that you can learn on this new behavior, as well as some articles regarding that stuff. Yeah.

Let me summarize this up with an article I've read and he put a nice conclusion into it. In short, the next time you craft a prompt, don't feel pressured to strip away the emojis, the pleasantries, or the extra content. Boomer prompts aren't relics of the past. They are tools for a richer, more intuitive future. So go ahead.

Michael (01:00:20.578)
I love this.

Ralf Richter (01:00:22.041)
It's really nice, yes. I also like it a lot. And it's true, in my opinion. Okay, Michael, there were tough discussions in our today's episode, a lot to take away. And we will feed you with all the links in our episode today. I do have something which I want to bring up now into Decode AI. The new thing is that for all of your...

out there who are developing software with AI or who are going to start with that. I want from now on to bring always a tiny tool or two or maybe two, two or three tools with me to explain to you because I really like that we're not only talking about AI stuff and politics and the behavior of it,

It's also important to me to give you some takeaways with you when you're looking forward to develop your solution. And on my opinion, the very first thing you want to have with your AI when it comes to something which is running within the EU, you will have to consider the GDPR as well as the EU AI Act.

And thus, you may want some tools to assist you with that. And the very first thing I want to share today with you is a tool which helps you to anonymize, or how we call it in the AI developing world, to mask data which is considered P2 data, so personal data or...

confidential data which shouldn't be exposed to your model or to anything else there. And you can reach that easily by a tool which is conducted by Microsoft called Presidio. And Presidio is capable as a tool to give you the possibility to

Ralf Richter (01:02:45.617)
find and analyze your text, you're going to process with AI. And after you have identified it, you also can anonymize that. And that's not the only thing it can do. It also has the possibility to redact images. So that means images who are considered to contain

confidential data can also be processed by a procedure to be masked and then being processed with AI. Whenever there is a concern that the data is highly confidential, you don't want to share with your LLM and thus with maybe a hyperscaler, you can utilize a procedure for that. And it's

capable to do that on many different ways. And if you want to explore that, you simply go to hugging face. And on hugging face, you can select Presidio. And there you have the possibility to explore Presidio on a demo where you can see on how it is conducting with its different approaches. And one approach is just by finding

entity types like a person, location, credit card information, crypto information, email addresses, e-ban codes, IP addresses, and so on and so forth, So with that, it is pretty nice and easy to start with, but you can also go for regx expressions, for instance.

or rule-based recognitions and so on and so forth. So it's really a pretty well-in-place tool to get started with taking care of your data, to mask your data, to be processed with AI. Have you ever heard of it before, Michael?

Michael (01:05:07.546)
That's kind of cheating. Yes, I heard about this last week where you are talking about this at Global AI Chapter Munich meetup. So yes, I heard about that before, but before that, I haven't heard about that. And I really like the idea to put in some not anonymous data with personal data, and then you just can anonymize it.

And I just clicked the demo from hugging space, that right? Hugging face. And it's impressive to see how this affects the data. Also, you can have different thresholds. So that's something you can use to eliminate data you don't want to have.

Ralf Richter (01:05:42.597)
face, yeah.

Michael (01:06:05.624)
but also you can keep some data if you like to keep some specific data into the data you want to feed to another system. And that's good. I like the idea. And thank you for sharing that.

Ralf Richter (01:06:13.357)
Yeah.

Ralf Richter (01:06:21.413)
Yeah, thank you. So it is not only that you can mask data, you also can hash data and so on and so forth. So it helps you by the first steps when it comes to, let's say, Rack implementations or so. is really a good and nice tool. It can run on-prem as well. So if you're thinking of something hybrid, it may be very handy for you. Great. So you brought up one important thing, Michael.

meetups. What's that?

Michael (01:06:56.45)
Well, this is where you can meet people and it's not a kind of blind date. Usually you can meet people with a specific topic or yeah, some people with the same interest. So meet up.

Ralf Richter (01:07:09.649)
interest.

Michael (01:07:20.372)
Usually there's a kind of platform where you can use meetup.com for looking for specific meetups in your close area or if you're visiting another city, you can search for them and some are free, some you have to pay for it. But this is a good platform to get some ideas about what's going on for specific topics without knowing much about...

connections or something like that. It's free. So you can search on the platform for free. And yeah, we, you and I and Thomas, another colleague of us, are running one meetup together, which is the global AI community in Munich. And we used to have about every second month.

Ralf Richter (01:08:19.707)
We try to be monthly, but it is tough due to some dates and appointments and so on and so forth. Yeah. We try to be as close as possible to being monthly. at least, these days, it will happen like being monthly because we had it on the 10th of or 11th of February and we're going to have it now on the 10th of March, right?

Michael (01:08:49.006)
Yes, the 10s. And we talked about the 11s with a hands-on lab for agents, And yeah, workshop, workshop. That's what I meant with hands-on, but it's more precise, exactly, workshop. So you can build it on your own. You're not allowed. You should build it on your PC.

Ralf Richter (01:08:58.875)
workshop.

Ralf Richter (01:09:04.155)
Hands on.

Michael (01:09:18.538)
with some descriptions, but you can make your own experiences. I just want to raise the attention to another platform if you're interested in some AI topics, maybe you are related to Microsoft, then you can also go to globalai.community.

which is a global platform. You can find different chapters also beside the Meetup. And you can find local chapters where you can have, where you will have different topics. It's not always talking about the same topic. It's not aligned or something like that. It's about talking about AI.

Michael (01:10:10.872)
So yeah, that's basically some kind of, well, that's meetup on a very basic level and very specific our global AI Munich.

Ralf Richter (01:10:23.205)
chapter. And we're running also, is it on what date is that? So the 10th will be virtual, right?

Michael (01:10:31.47)
The 10s will be virtual and the 11s will be on site in Munich.

Ralf Richter (01:10:39.951)
Yeah, so and it'll be a workshop where you can bring in your laptop and everything to follow the workshop and to create it on your own. There will be some prerequisites, but we're going to share that with you before.

Michael (01:11:00.246)
Yes, you can find the link to the meetup and also to the global AI community in the show notes, of course. And then you can have a look. Also, if there's maybe another chapter closer to you, feel free. Just go ahead and give it a chance maybe to join Karlsruhe, Berlin, Bochum, London, Paris.

Tokyo and many other cities. So it's really globally.

Ralf Richter (01:11:31.259)
What the?

Ralf Richter (01:11:34.607)
Yeah, that's cool. And what we will share with our show notes as well is the Experts Live happening this year in Leipzig. And you can claim a voucher to reduce the ticket price, which is dedicated to the communities.

Okay, great. Michael, do we have something to add at least?

Michael (01:12:06.686)
No, I just want to try the catchphrase. What is it? The closing catchphrase?

Ralf Richter (01:12:13.158)
Yeah

closing.

I don't know.

Michael (01:12:20.61)
So let me start with a thank you, for this episode. It was really interesting. I've learned a little bit or I learned a bit more and I hope you did as well. And also to the audience, if you have any questions, don't hesitate. Just reach out to us. We are on social media. I think LinkedIn is the most preferred platform. So, and that's...

Ralf Richter (01:12:25.649)
Thank you, Michael.

Ralf Richter (01:12:33.647)
Yes, for sure.

Michael (01:12:49.378)
the point where I would like to close our actual recording. And stay tuned, stay interested, sign up, listen up. Here we go, bye bye, take care all. Thanks for listening. See you next time.

Ralf Richter (01:13:03.601)
Thank you, bye bye, take care.