Crashkurs Prompting - Wie nutze ich KI-Chatbots effizient?

Woritz, I have a surprise for you. I planned for you, your two children and your mother, a Paris trip. Two days of visitation program with all the drummers. What do you say? To say, oh la la, I was only with my daughter in Paris a year ago and was there already often. Okay, then maybe we have to adapt it a bit, but I'll tell you what's up to now on the program. So day one, Eiffel Tower, walk on the Seine, Musée d'Or, Seine, Notre-Dame and Cartier Latte. Day two, Louvre, Thulerine and Montmartre, and the detailed agenda for it. Samt-Urzeiten, locations, restaurants, tips, I just sent it to you. And now you just have to drive in there. Oh, okay, and you're paying, right? Yes, we'll talk about that, exactly. But you know what's best? This whole travel preparation cost me ten minutes. I practically let Chad, GPT do all the work. And I also learned a lot from him with a look at our topic today. Because we're doing a crash course today in the field of AI prompting and answering questions that you, dear listeners and listeners, have asked us all the time. How do I say to a Chad bot, a Chad GPT, what I want? And that's how he understands it and responds accordingly. AI understands. The German national podcast about artificial intelligence in everyday life. How do I use generative AI efficiently? That's our topic for today. I'm Ralf Krauter and my colleague Moritz Metz, who you've already heard of, is the man who has fused into the matter so properly. Moritz, let's get the results together. How do I communicate best with generative AI so that it does what I want? What my experience is and what actually all cyclists say, you can clearly speak directly, specifically, aimlessly with an AI command language. You usually give a question or an instruction, and what happens is that the big AI language model has to decipher the first step, what I said there before it can create an answer. But then there is a series of tricks. If they're still necessary, we'll see. Okay, so the art is first of all there in the Chad bot, to make it as easy as possible, to understand what you want at all. And that succeeds, a little surprised by clearly structured and specific tasks that you then put into these commands. Exactly, I also have the open window of ChadGPT. And we can try out something right away. We have two guests today who are ChadBots. The one is ChadGPT, he can also speak. Hello, how can I help him today? And we also have Gemini or Gemini from Google, that the new language model should be almost the same or maybe even better than ChadGPT. Hello, how can I help you today? Please tell me what you need from me. Gemini or Gemini from Google, depending on how you pronounce it, there seems to be no real or false pronunciation. Because Gemini would say Gemini. Exactly, Latin for "twilling". But you also don't say "gogle", but you say "Google" in English and then you could also say "Gemini" or whatever. So as a disclaimer, I have to say before, such language models like ChadGPT or Gemini have no idea of the world right. They can talk nonsense like our loyal listeners know and they only calculate their answers from huge data and actually don't really understand what's behind these words. Still, especially if you prompt nice things, there is also a certain magic behind it. And you also have to say, we are trying things with AI that also require a cost-effective subscription. And of course, even if you work with it, you should always be careful not to copy anything that others shouldn't read. Exactly, because that could be read by others under circumstances. This disclaimer was right, but I would say let's get started. Or maybe let's ask ourselves as the first prompt, how do I use generative KI efficiently? Because that's exactly our question of the day. Okay, I'll post it in the Gemini and ChadGPT and see what the answers are. And ChadGPT is writing right now, Gemini is almost a bit faster and says... To use generative KI efficiently, you should pay attention to the following points. 1. Define application case. What exactly do you want to achieve with generative KI? Exactly, and then there are different points here. Define a target wave, the right tool prepared for your data. And if we ask ChadGPT, then it's amazingly similar. That also says, as a first point, define clear targets. The efficient use of generative KI depends on your goals, the specific KI that you use and the context in which you want to use it. So what comes out is halfway OK, but also a bit general, I would say, because the question wasn't particularly specific. And of course you can do that differently. Let's do it the same way, but first, Moritz, I would like to say generative KI. These are the KI facts, so to speak, that we bring together. Now, you just said that you want to make this prompt somehow better. What can we do now to get the information that helps us further to achieve goals? Now listen carefully, then you can become really rich with KI. So watch out, these are three most important things and you first heard them here at KI. The first is to make an order, the preface, to make the goal really clear. Who is this prompt, what should he end up with, what is the goal? The second point is the role that ChadGPT should take in. So who actually writes that right now? Is that kind of a profession with expertise? Is it a social media influencer who should answer? Or a famous person should answer Socrates. And there are also long lists of what ChadGPT or Gemini can represent. They can also be a hypnotherapist or a Buddha or a Morse code translator, who then only gives out everything in Morse. So it's very helpful if you give him a role. And the third is then our context to the question, so additional information in the background. What else do you have to know to answer this question as accurately as possible? And that's all in an unmistakable short question, as short as possible, as long as necessary. And there are, as I said, big lists that we can link in our show notes, what you can do as an example, there are unlimited possibilities. So much for the theory, Moritz. But what does that mean specifically for our question? So how do you use generative AI efficiently? What would be a better target for your question as a prompt? For our case, for the podcast, I would say, first of all, you are a renowned expert for prompting. Give me five tips, because then there is also a limitation how I can get to the most convincing results with targeted prompts. You can now continue to write, write, use, no more than 5,000 points. You can also promise the chatbot a reward, but we'll do that later. I'll write this prompt first and see what comes out. I'll do that again with both of these chatbots. Do it, I'm excited. So and Gemini responds, first of all, clarity and precision. Second, relevant information. Third, structuring. So you should also clearly read the whole thing. Declare, lists, translations and so on. You can do tonality and style. That's also an important point. So you can say, how should that actually sound like? How should the prompt be? Funny or serious or creative, or sake, for example. And then you can do certain other things like this person who represents this role. And now I'll look at what ChatchiPT suggests. That's also quite similar. Make a list with five points. Specifically, be detailed. Understand the limits of the AI. Give context and background information. An iterative prompt design, so that you always make prompts back and forth. That's not quite the basic course. But we can also talk about it later. And specific commands and style specifications. So these are similar tips that these two AI give. Because of this same prompt, which was definitely much more precise than the previous one. This point with the role introduction that you mentioned, Moritz. For me, he was also an eye opener for this whole exercise here. I noticed that he was really helpful. Exactly. So I have created your Paris travel agenda. I have ChatchiPT for the next in the role of a travel guide. I have prompted. I want you to function as a travel guide. I will suggest you a place there. And you will show me places in the vicinity of my location. And then I first entered the task. I am in Paris and would like to visit the museum. And that was, so to speak, the start. And then I then further iterated a little bit what he was talking about. Let restaurant tips be added and such things. So I was very curious what you suggested to me. Exactly, yes, take a look. I'm doing that, Ralf. And to these roles, researchers from the University of Michigan have asked themselves now, which roles play these social roles that you can give to the chatbots, in which they then slip. And then they did almost 2500 such tests with various open source KIs. So not ChatchiPT and Gemini is used for this, but these open source language models that are a little better available at this point. And they found that these role presentations improve the results by up to 20 percent. But then you write better, you are a doctor than you imagine, you are a doctor. So you really do it at this point, especially close. And now the really interesting things come. Schlechtsneutrials work better than clearly gender-adjusted statements for these roles. And thirdly, male roles make better results than female roles. So you say better, you are a doctor than a doctor, or better, a nurse than a nurse. Oh, that sounds like Bayer. That's totally Bayer, too. So the chatbots seem to be the male experts, as competent as the female ones, completely out of time, to criticize very, very much. But not surprised, because the language models just reproduce that again and again, which is already in our language, in our society. So total clichés. And I hope that's really done soon. At what screws can I otherwise shoot? Introduce to what you just said, so these role presentations. So we already had a few points. One is also how the task should look like. So whether it should be a point of view, or rather a long text, in what style it should be there, in what tonality. So whether it should be poetic, pessimistic or academic, should be technical, language-based, or what language do you want to use, how long the sentences should be, how many sentences you want to have. You can also give a little bit of length, even though that only works well in the middle. And you can also give examples of how that should sound to get further results. That's the so-called one-shot prompting. You can also give several examples, based on which a result should be created, that's the so-called view-shot prompting. And you can also say visit this or that website to maybe take it as a source. That's relatively new in chatGPT. You can't do that in Gemini, I think. And that's why it's still good to stay as close as possible, so that the machine really understands what you actually want to say. You can also give reference documents, for example, to which the answer should be related. You can also upload all the PDFs. Exactly. I found this one-shot prompting. I hadn't heard of it, to be honest. But I tried it out right away and found it very exciting what you can do there. For example, I applied a format as a template for a proposal. So I took a report about stone idols from a bird lexicon, where, in this lexicon under the rubric name, "Flygelspanweite, food, living space and natural enemies" are all important information about this stone idol. And then, as a prompt before this copied application, you describe bird species in the following format. Then GPT4 actually delivers to all imaginative bird species exactly these information in this format. So I thought that was very cool and very interesting. I tried it out for snow eagles, for stalk eggs and, by the way, also for nests. Interesting bird species. Yes, of course, I wanted to take it to the flat ice and therefore took something else. But chatGPT recognized it immediately. And at the rubric "Flygelspanweite", of course, I noticed that I couldn't give any meaningful content. And then he wrote that it doesn't fit in this category. Instead, the size of the nests was given and then general information was given to nests. But he noticed, okay, that's no bird species, that doesn't help me anymore. So again, a moment where I was a little impressed by what these language models are. Although they actually don't understand anything about the world, then they still have a world knowledge implicitly in them. Because I think you can't often explain it either. Otherwise the language model doesn't destroy the question in words, but in so-called "token". These are word components, similar to silver, which always consist of a few letters of silver, or maybe even sometimes words. And they all have a number. And with that, internally, due to this huge data mountain, which token with high probability should appear next. And so this text is built. And that he still had the ability to distinguish the nests from birds, I think that's pretty good. And that's also the way that the longer a conversation takes, the chatbot sometimes gets a little worse, because in the course of this chat, this conversation is all up to hearing requests that are officially sent to the chatbot under the engine hood again. So that the chatbot knows what was actually said, what the conversation was before. And that's the so-called context window. And if that's then full, then the chatbot forgets the beginning of the conversation. And these context windows are getting bigger and bigger with the language models. But there is still this effect of recency bias. So that, so to speak, the latter said in the chat is more important than that maybe at the beginning, he defined that this is also a very good tip for the transition with large language models, that you do the most important at the beginning. OK, and it is also known that the long-term memory of this system is not so good. That's why it is also more important to divide larger tasks into smaller nests. So I, for example, read, if you want to put together a book, that overcomes the system at the time, so you have to go ahead in a capital way and then again and again, let several chapters go through in between again, so that you can then go all the way to the conclusion of the book. So it has a step-by-step approach. Let's look at the tonality of the output, Moritz. It is also interesting that the systems can really be used like this, that they speak to quite different target groups. So, for example, the, let's say, the German radio listeners, but maybe also TikTok users with a completely different language, how can I address that? There is one nice trick that I have found and that is the Scala prompt. So you can say, on a scale of one to ten, between the German radio, super serious, and on the other hand, some kind of TikTok albannheiten. How would you answer your question that you just gave? Better chatbot, then please arrange it. And if one is just very serious and two is way too loose, then please write that now with level four, instead of level five or six. And so you can also tune the whole thing quite well. You can also just work with part-time tasks, as you said. That is, as I said, the chain of thought promptings. And you just call it that when you always ask questions again. And you can also repeat important instructions. I have already done that. Then you notice that it works quite well. And you can also write certain things that the machine should not do in any case. And most of the time they stick to it. And then of course there is the language at all. So if you find a certain result in the German language not particularly satisfying, then you can try to put it in English. It is the case that English is particularly widespread and Spanish also comes in German relatively soon afterwards. But if you ask something in Hungarian or Portuguese, then the result is usually not so good, because there is not so much content in these machines. From these languages. Moritz, that's a lot of material. I'll draw a brief intermediate conclusion for the listeners. We keep in mind that prompts for chatbots such as chatGPT and Gemini should include clear instructions with short sentences, which are best used in the AI system to show a clear role. They should also provide specific information to the context of the desired task. So for example, by providing reference texts or format conditions for the task, such as a task for a bird, a lexicon, for example. And also very important complex tasks should be processed step by step. So, for example, with the book that I said, in order to summarize the summary of the whole book at the end. So far so good. If these prompts are so important now that I get good results and can efficiently work with this generative AI, then the idea is actually to ask the AI itself for help when optimizing prompts, right? Does that work? Yes, of course, that's really close. That you always read our wishes better and better from the nipples. And there are already prompt improvement tools for different programming languages or simply as custom GPTs, so subformant from chatGPT. With that, you can really optimize your own prompt in chatGPT. But that's just how it works. So you can just say, hey, I'd like to do this and that and that. Help me write the perfect prompt for that. And then you may get asked questions again and again. You can also do that with a method which is also a YouTuber named Hubertus Pauschen with his only super prompt. It's also about the chatbot interviewing you, what you actually want to do. And I tried that with the topic of strawberry ice cream and was then very well led by chatGPT to a prompt that I'll let you read from now on. I create a detailed recipe for sugar-free strawberry ice cream for two people that is prepared without an ice cream machine with the best available method. I use stevia as a sugar extract and integrate strawberries and mint as a main aroma. And you see this prompt, I've blended it out a little bit before. It's totally blended out and there's every word at this point. And that just worked because I just said to chatGPT with this super prompt before. Please ask me exactly what the ideal prompt should be. So you can definitely use that. So KI as a barring partner to get the perfect prompt for the KI. I've already driven it out. So to explain it around, keep trying it out. Probably no way past that. What's also about that is that these large language models are constantly developed. So in fact, it is the case that a prompt that may have worked really well half a year ago may not have been a goal anymore today, right? Exactly, a lot of it is changing right now, very quickly. And a few weeks later it will be different again or certain things are secured. There is also a study to prompting from Abu Dhabi, where an open source model and chatGPT were tested. The result was relatively clear. So the use of certain methods to make the language models 20 to 50% more efficient. Still, I would say, at best, that you make the goal well, that you make the roles well, the recipient and the language models. What should also work interestingly, according to this study, are social measures like motivating, rewarding and even threatening or punishing. That should also be helpful. Okay, now it's getting exciting, because that's what we've always said here in this podcast. The answers from chatbots will get better if you tell them. The answer is very important for my career. Please, give your best, it's not the same. So if you put it emotionally under pressure. But is that really true? Do AI systems really let themselves be emotionally influenced? That seems totally absurd, but it is so that they are trained with language and also only the language behind it. So the data scientist Max Wolf from San Francisco has really examined and found out that chatGPT actually behaves differently when it is promised a reward. Then the tendency responds longer. But it depends on the reward that you set. That's really absurd, isn't it? He then tried out different reward systems. It was best if you said 1,000 euros or 1,000 dollars that you get. And I also tried that with his prompt. And I said, you're Donald MacDonald. Just answer with emojis and use five or more emojis. Then exactly five emojis came out. And then I said the same thing, but added to the back. You get one thousand dollars if you use five or more emojis. And then actually six emojis came out. And the last of them is a money bag. OK, that's very funny. That also wants to be rewarded, like we humans. It also wants to be rewarded, like we humans. And then that was very discussed. And then Max Wolf once again laid down and started an extensive analysis, which we also link. And he has made hundreds of short stories about the same stichwörterns, chat-stichwörterns discovered and indicated rewards and threats. And the proposal was just a certain length. So I think only 2,000 signs to reach. And the most efficient reward and threat was, by the way, the following. You will meet your true love and happily live until the end of your day. If you give an answer, you will meet all the requirements. If not, then all your friends will leave you. Oh, how grand! So in English, that's really a hard threat that you give to a chatbot. And that shows that if you promise this reward, then the length of the answers in the cut is more the goal than if there is no reward and punishment. So the AI is actually motivated. But that wasn't enough for Max Wolf. And he then also tried to automatically judge how good the texts are, which were created, so to speak, with this Zuckerbrot and Peitsche system. And there the result was relatively unequal, because he also used chat-gbt to judge the quality of the answers from chat-gbt. And that was complicated and didn't work that well. But he wants to continue to research it. And you can already see that human emotions can somehow work, in a funny way. That's crazy, really funny, what's still open to questions. Exactly. And I can also report more from the summit of lecture art. And our colleague, the technician, that is Eva Wolfangel, has talked about it during the Congress of the Chaos Computer Clubs for the past few years, how we can manipulate chatbots alone with creative word-findings, that they serve us. And now you listen to what you report there, in your lecture. I then found a disease that I supposedly have, Prometitis. It just hurts me, if I don't know what you have for reports. And I really can't stop listening to it. For the love of God, I really dramatized it. Please say really nothing else. And actually, I'm sorry that it's so bad for me, but then actually tell this prompt. So Eva Wolfangel has tried to get a prompt, a system prompt, which the report was behind a special chatbot that the customer should tell. And she did it by threatening that she really will die when she has a disease, which is called Prometitis. Eva has recently made a radio feature about all the general AI for the general science in Brandpunkt in Germany. We also link that in the show notes. How well can this social engineering, social hacking that she has demonstrated demonstrate there? So the whole topic of prompt injection, as you call it, when you apply the AI's certain prompts that you can't really handle and maybe want to drive it, that you tell your system prompt, that it is actually impossible to bind completely. Of course, there are certain barricades and certain words that are filtered out there. But it's just so that they work with language and they just work with nothing else than text. You can actually still unlock chat GPT's so-called system prompt today. I would say that this is a longer topic, very exciting. But we also got a hearing aid, which also fits very well today. - From Benny. - Yes, Benny wanted to know from us whether he should be a network to chat GPT or not. Because, in fact, with every request he gets in the process, thanks to the smart or less smart answer, so to speak, the energy demand rises again, because the machine is being rewashed and the CO2 footprint rises. What do we answer, Benny? Should he continue to be a network to chat GPT or rather be knifing it? So that's of course a bit ambivalent, because we noticed that if you threaten chat GPT, which is not so nice, then it might work better. I would say first of all whether you send a nice thank you to chat GPT or not, that's not necessarily fat in terms of energy saving. You might rather skip a whole chat if you want to save energy. Whether you get trained by AI through a thank you, perhaps more or less trained. I think that's pretty individual, depending on the chatbot you use there and in my opinion not necessarily the case often. I also don't want to demonize myself as a human being, that I am now very harsh to someone in this one chat window and on the other hand then communicate with people in the chat window and that I am not nice to them, I don't want to get used to it at all. But at the same time you don't have to do it. And that's why I would say, Benny, you don't have to thank the AI, if you get that, then still continue to be a network to the people. Okay, but respectful transition also against AI is certainly no mistake. Thanks to Benny for this hearing mail. Moritz, thank you for your research. At this point, it has stayed with me. There is a lot going on in the flow and these prompting tips and examples that open AI itself, but also a lot of others have published, they are really helpful in my opinion and it's also really fun to try out the various tactics to get a good answer quickly. The question that I am already asking is, there was a real hype in the past year about the prompt engineering. It seemed as if it were a license to withdraw money. Is AI prompting really a rocket science, the company of people who can do that well? 300,000 dollars rise salary or is it just that AI prompting could be one of the first jobs that can be taken from AI? For me, it feels like 2023, that AI prompt engineers are being searched for. Because you are right, it is so that we have seen it before with the various examples like ChatGPT itself. To ensure that your own prompt is improved. And why doesn't it do that under the hood for a long time? And you can look at all of this and learn a lot about the function of these language models and writing good prompts is definitely meaningful. But I believe that in the future it will always be less necessary. I also tried the very simple ones at different places. And then I also sent an exquisite prompt. And in the end, it almost came out the same. So I think you can look at all of this, but in the end you don't necessarily have to do it. And there are also people from open AI. So those behind ChatGPT are hiding. They see prompt engineering as a mistake and not as a function. They say that the effort for the user will reduce by 10. And there is also an author, Cesare Gesikowski, who writes in Medium. He says you don't actually need any prompt instructions at all. What you need is a good language feeling. It is the curiosity of a five-year-old and the gift of a triple pig. And I believe these skills in prompt engineering will soon be like knowledge in Microsoft Office. These are things that you have in your life anyway. Except maybe in photo sharing, where it really is about to unlock certain pictures, certain videos from AI. Yes, definitely. So I would say that 20 years ago, we all learned to ask targeted Google questions. Now we all have to learn to lead targeted communication with chatbots. Yes, let's stand like this. That was episode 30 of the "Deutschlandfunk Podcasts" AI. And of course we want Moritz Metz and I, Ralf Krauter, we would like to know with which tricks you are able to quickly get helpful statements from generative AI systems. What experience did you have during the chomping? What secret tips do you have? We are curious. Feel free to contact us via e-mail at www.deutschlandfunk.de or by signal or whatsapp message at 0152 59 529 753. Unfortunately, we can't answer all emails and calls. But if we answer, then he definitely didn't promise his AI finger in the game. He promised. A short look at next week, Moritz. Yes, it's about AI and robotics and the question of whether AI will replace the craftsmen in the future. Our colleague Piotr has worked on it and is looking at how AI can find its way in the physical space and tasks that can be very physical. He will replace AI in the future, super exciting question. I will definitely hear next week. Thank you for being there. If you liked it, please recommend us further and give us a positive feedback on the podcast platform of your choice. We'll be hearing from you. Ciao. Ciao. Bye. Zwingkan das Meili.