Welcome to the IT Communication Research Institute, Director Kim Deokjin. Hello, nice to meet you. Oh, a new book has just been released. So, back then, I talked about AI 2024, and I was actually preparing for AI 2025. Yes, but there were significant changes in the middle of it, and I also thought that the chatbot market would become a big industry to talk about separately. My book only talks about Bookbon 2025. It only talks about Bookbon. But when I say chatbot, we used to think of things like customer service consultations. But now, it's an era where everyone has their own creation, right? You can combine the book I made with the book someone else made, and now it's not just about using your own book to end with. In the future, this will be monetized.

I think it will be connected to Azure, so if someone uses the punishment I made, they can make a profit as much as they write, and also by utilizing various functions and hidden functions of this chatbot that were not well done in the past, it can be like this. For example, you can add things that are unsatisfactory to chat GPT man additionally, such as attaching someone else's file or connecting to external services to structure this chatbot. In the past, it was really difficult to create such things, and for example, it took a few months to attach to a professional company, and it cost tens of millions of won. But now, if you just join GPT, it costs about $22 a month, and you can make it fun, and no matter how much you look at it, this chatbot, Chet PT, is still being significantly developed into a platform.

So, because we now have the App Store, in a way, the iPhone has become quite superior. By creating a chatbot with just text like this and sharing it in the GPT space, it is gaining popularity. As time goes by, when it starts to generate revenue, perhaps in the form of in-app purchases, similar to apps, it might expand. Because of these factors, it could be easier for you to use and follow along. Like last time with AI 2024, this time we are making a chatbot. Even if you don't make one, there are already a lot of templates available. Making them now is a bit more challenging than before, so there are really a lot of ideas. Therefore, I recommend trying out these templates and also showing you how to follow along.

Even if you can't do anything, I want to try making something like this once. Then we have a prompt lined up, so you can just write following that, something like that. I've also written about 40 of them today, so soon I'll ask you to do a little Zion. It would be good if you could check how simple or difficult it is to make a book. And most recently, there is Chechi PT 4 Omni, Chechi PT 4, and 5, which are now in the air. Expressing emotions, and later, through vision, using a camera like a phone to see and speak on the screen, and even have a conversation. While watching GPT 45, what was the most surprising point for you? There are two things for now. One is expressing emotions, and this expressing emotions hasn't been given to us yet in terms of recording standards. So now,

When you try the voice function, it often comes out in chat that something like that will be available soon, but something quite meaningful is that it becomes a conversation. What does that mean? In the past, we could use GPT for voice recognition, but now we can use it for voice chat. However, when I watch various YouTube videos or broadcasts, it feels a bit like walkie-talkies. What I mean is that I have to listen to it all when I start a conversation and it responds. Yes, that's right. That's how the story unfolds. However, now people can interrupt the conversation in the middle. But if the voice recognition of GPT-45 is properly integrated into our phones, based on the demo, you can interrupt in the middle. For example, if I ask something and make a strange sound, then say, Stop, stop, don't say that, say it like this, then it will respond in the middle or say, Hey, you're so boring.

When you talk, try to express emotions more intensely. Then, like talking about ancient predictions, and then, as you mentioned, multimodal has become perfect. So, in the past, there were separate actions for taking images, having voice conversations, and writing. As you said, now when we take pictures or videos, we really can have conversations while taking them. It's like having an agent next to us, like a kind of assistant or model, as you mentioned, that can be with us completely. It's something that can exist now, although it hasn't been fully introduced yet. Watching the demo was surprising, and writing and being surprised from a business perspective, the speed has become incredibly fast. So, those who have tried using GPT-4 would understand, but GPT-3 was a bit frustrating. However, then this character Claude appeared in the middle. Claude, this guy, appeared in the middle.

Former employees of OpenAI have now created a kind of arrest service separately as entrepreneurs, and it was quite fast. So at first, when they used Claude, they said, Oh, we can't use GPT-4 anymore. But GPT-45 is even faster. It has become much faster. So, while using that, when they tried to use Claude again, they thought, Why bother with Claude? to the extent that people are a bit cunning. The second point is that this is now somewhat connected to the arrest, but it is definitely solidifying OpenAI's business strategy. So, they have now clearly started a kind of partial paid and strategic strategy. This means that most likely, those who are interested in watching videos have already become members of Cheachi PT. One of the biggest changes in OpenAI recently is Cheachi PT.

You can use it without signing up. Without signing up, in a state of not logged in, you can just use it like typing in a search engine. So what does this mean? The hurdle has been lowered. So, anyone can just try it out without signing up. Even when I actually go to teach, what was the most difficult thing for me in the past was that I had to give a two-hour lecture or a one-hour lecture and then do a demonstration. Signing up takes away 10 minutes. Because, surprisingly, you need to go through phone number verification and email integration, so people tend to think it's quite a hassle. But now, you can just use it. Just use it, so you can get a taste of it. So, there is version 3.5, right? There is GPT 4, 5, and then 4, and there is version 3.5. From 4 onwards, it is a paid service. Like you mentioned, before signing up, it's 3.5.

Criteria, but if you just sign up from this time, the changed GPT 45 can be used for free to some extent as a trial. Yes, this is a huge deal. So, what is it? The speed and feeling of GPT 4, these kinds of things. Now, if you use it several times, then you need to wait or sign up for a fee. This is how it works. In the past, you couldn't use the latest technology without signing up. Oh, so in the past, if you didn't pay for the latest technology, you couldn't use it. But now, everyone is using it. Therefore, within this GPT, there are GPTs, or GPTs, and even the book market will inevitably grow. So, before GPT 45 came out, only paid members of GPT 4 could use this or utilize it. Right? But now, because GPT 45 has been partially opened to free users, it is being applied sequentially.

However, the arrest I have now created can be used by free users if they log in and simply sign up. Of course, to create it, you have to subscribe, and then you can use it, which is a significant change. For example, when we make a waistlet with Director Kim Deok-jin today, we naturally signed up and created it as paid members, but even those who view it for free can access the waistlet BT 4 and use the service. Yes, that's correct. Now, this is being applied sequentially, but anyway, after signing up, after a certain amount of time passes, or randomly, or as the application continues to increase, probably within a month or two, anyone, including free users, will be able to use the service. However, here, in terms of the business model, yes, there is a limit on the number of uses, so there is a limit on the number of conversations.

So, when I have a conversation with GPT-4 or 5, or when I talk to a chatbot, there are limits on the number of times per hour. So, if you think about it, Hmm, this feature seems nice. I want to keep using it. There are people like that. They used it up to ten times, but if they want to use it again, they have to wait two hours. However, instead of that, we can naturally draw a picture of revenue sharing for membership to chat, or for each conversation, or for a 30-time conversation coupon, or for paid models like this. But I think we should also talk about that. Many people think, Oh, that's necessary. It's just a search. Nowadays, searches come out in detail, and even chatbots like Zeplin or Naive speak in a conversational tone, pulling out everything. Why do we need to use or create a chatbot? First of all, there are three main reasons.

I think you can consider it as a level of GT. We always need to set rules of commands in the prompt. You don't always have to do those things. What does that mean? It means, as you mentioned, it would be nice to try this later. If we have a book with the editing style of Lee Dae-ho, it wouldn't be good, right? It wouldn't be good. But how should we do it if we want to create that now? Of course, if you ask GPD 4 naturally, it won't come out in the tone of Editor Lee Dae-ho, right? I don't know, sorry. Well, then, there is a way to do that. It's called prompt engineering. So, in the context of the conversation I'm having, like, From now on, you are Editor Lee Dae-ho. Editor Lee Dae-ho speaks like this and is this kind of person. Okay, let's start the conversation now.

If you understand the context of the conversation, but the important thing is, if you open a new chat or talk about something else in the middle, it forgets again. Oh, when you turn off the browser and reopen it or chat, or if suddenly, instead of 2 to 5 reporters in the context of the conversation, someone else's content is brought up, it gets all mixed up. For example, suddenly, instead of 2 to 5 reporters, Kim Seokjin and some strange 2 to 5 get mixed up, some strange characters can come out. But I can give these settings in advance to the chatbot. So, for example, these things can happen. I will show you later, but for example, I create Warren Buffett, and I create Donnamu Unni. Then, I make a book of Warren Buffett and Donnamu Unni having a conversation. It would be fun, right? And then I give the content, so the characters can talk.

It's about being able to tune a chatbot the way I want it to be in this form. That's the first thing. However, what's done there is just tuning within the information that GPT already has. Changing the tone, changing the context, giving a character, things like that, so the content itself doesn't change up to that point. But now, the content doesn't change, but instead, I can continuously change the tone, feeling, and such of the content the way I want. That's the biggest thing. Recently, as you keep using GPT, you talk a lot about it. Oh, it's good, but it's too inaccurate about our company. Or maybe, I think I'm better at this kind of information, but it doesn't come out like this. So, it's like...

I am very interested in those who say they would like to study additional information about our company's information or the information I provide and answer based on that. Well, that's possible. What do you mean? For example, in the choreography I created, you can upload a PDF file or provide some information, like giving the content of my book, study the content of the book, and have a conversation. So, when making a chatbot, it's difficult to properly refine the data and it's also very important to speak well about that data. You should think about that once. Banks and oil companies also have services like that. If we ask accurately, they provide good answers. For example, if we have a problem with a public certificate, how to solve the issue of a public certificate, they provide good answers. But if I get angry and say, Why is this happening? This public certificate is a mess, and so on.

Just by talking like this, if the context gets longer, the existing pattern is that they might not understand it properly. That's why we don't use the term arrest much and prefer the term physical therapy. Anyway, whether you ask this kind of question or that kind of question, we answer, so it's okay. But the problem is, this person doesn't have accurate knowledge about the procedure for resolving the public certification of a certain bank, for example. They don't know how to reissue it or renew it, but they are aware of the recognition. If I can't access it now because of the public certification, then the problem is that the child will be disappointed. Right, because the procedures can vary from bank to bank. And even though it's the same question handled by company A, for example, the way to press the menu in bank A and in bank B can be different.

Also, that menu could keep changing, right? So, what do we do then? GPTS provides a manual-like PDF file or gives other additional information. That is, providing information or knowledge. That's how it is. A manual unique to Bank A, a manual unique to Bank B, those kinds of things. Then, when we provide that, the person will study it in a way like looking at the key points of a book or a PDF file I gave them right before answering. It's like looking at a test book before taking an exam. So, when they answer, they are friends who are good at speaking originally. For example, there could be people like this, right? Like shopping hosts or show hosts, who are really good at speaking. They are good at speaking, but someone else precisely inputs the content. Right? They are given the key content right before the broadcast. That's how they speak.

If you do well, by looking at the content, you will respond well. But originally, to create such a solution was very expensive. However, with GPT-3, it's possible. But the companies that develop these programs used to make chatbots and earn huge profits. If this comes out, it will greatly impact the chatbot developers. So now, there are many concerns like this. But as I mentioned in my book, the approach is a bit different. How is it different? When we used to create chatbots, it was like making a sculpture. There are two ways to make a sculpture, right? One is that I continue to add intelligence and things like that, or keep adding them one by one. This is the traditional way of creating. So, by adding precise queries and precise content one by one, it is made. When it's done like this, what happens? The accuracy of the arrest is high. But with GPT-based...

The opposite of a choreography is cutting. There is a chunk of cutting, and that chunk was given a document hastily. But for example, if someone suddenly gives me too many documents all of a sudden or I give them in a rush, just five minutes before the live broadcast, someone who speaks well suddenly comes in as a substitute, suddenly giving me a 30-page problem set. But even though I glanced at it roughly, I got it wrong. If that happens, after the broadcast ends, for example, I need to say, Oh, this content is not correct, it should be fixed like this, and then in the next broadcast, I won't make that mistake. So, this kind of continuous process is necessary. Therefore, in creating GPT, the method of creating a chunk like this first and then refining it continuously is a bit different. Next, from a personal perspective, this is quite good, in the existing business, if

When you calculate the cost of a typical arrest and this arrest, it seems that now we are creating it within the book GPT, right? So, when creating within the book GPT, we pay a membership fee of $22 per month, and then we don't have to pay additional money. However, if a company wants to create something like this, they need to use an API, right? Using an API means bringing in the GPT engine. After bringing it in, you attach a file there to make it conversational. What does that mean? The more people speak, the more money they make. It's OpenAI, right? So, the more questions people ask, the more they pay per question. Therefore, this changes the structure of the business. Individuals can just use it for $2, but if a company wants to create professional content, OpenAI charges more. So, in what way is that, for example?

If a company that knows about the service wants to add this open AI chatbot service to their service, they attach it to the book GPT API service. Then, through that, they enable conversations. Like I mentioned earlier, they provide company documents to study and create a service where people can have conversations to correct mistakes. By doing this, they say, We are ready. Then, every time people use it, the conversation grows because they ask, How much is this? How much is that? So, the number of conversations increases significantly. Consequently, the operating costs will be high. Therefore, ultimately, it is a matter of choice. So, I want to say that recently, I have been talking about the distribution business of generative AI-based, especially GPT-based API businesses. So, from a corporate perspective, it is quite different from a personal perspective. From a personal perspective, it's like, I can write such a good post and use my information for various purposes.

Since this is technically oriented, you need to write quickly and write a lot. However, from the company's perspective, when actually introducing this to our company, it's a completely different story. As mentioned earlier, if people use this, for example, more than 1,000 times, the amount will be this much, so intentionally lowering the version is also done. Therefore, the latest version is expensive, and as the score goes down to 3, the price per company decreases for each question, so we also calculate these things and this is how the content is created. In a little while, we will try creating a waste street cell once, and since our subscribers are approaching nearly 500,000, if many people use it, we could be in trouble. So today, what we will do is create it from individual accounts and only in a way that does not cost money from individual accounts. It could be a big problem if done through a company account, right?

Yes, instead, what's interesting about using GPT like this is that individuals claim to be having a conversation. So, they are using their own conversation data. What does this mean? It was mentioned that GPT-4 cannot be used unlimitedly like before. Therefore, if someone is arrested on the street and tries to have about 500 conversations in an hour, it won't work. Oh, I see, you set a limit. So, the limit is for the individual users, not for us. We are fine with it. However, I need to mention one more thing here. This is just providing personal information. Thirdly, today, I will just briefly show you that if you utilize the arrest in GPT-4 well, you can connect not only simple documents but also external services. This is quite powerful now, and it's more in the realm of development.

I need it, so after this discussion ends, I will show you that first. What I have created is, now, with Mr. Seo Seung-hwan, who is a co-author of a book with me, he is very good at programming and prompt engineering. I asked a friend who is good at that to analyze the content of a YouTube video. For example, if I just put a YouTube video link on the website, then it analyzes the content of that video. After the analysis, Warren Buffett and Donna Moo do a talk show together based on the analyzed content. With that huge amount of data, you can also create things like that. How does that work? Analyzing videos on YouTube is not within ChatGPT, so my friend created a separate service for that. But by connecting the separately created result, within ChatGPT...

It means that you can create it with a chatbot. So, what does this mean? Later, if I am really good at professional development, as a developer, I will create my own app. But if people don't use my app, then I will create a GPT script. So, I can create a service like asset management within the script. For example, I can run the content related to asset management on a server of a company and bring the results into GPT. I can not only use the results as a report but also change the tone to a 2nd to 5th-grade journalist tone. In this way, the service of asset analysis within the service can be transformed into a service where the results are spoken by GPT as if it were journalist Lee Dae-ho. In the past, if chatbot and chat functions were partially included in the application, now asset management services are completely integrated into the chatbot.

It changes, right? So, this allows you to combine various ideas like assembling Lego blocks. What's even more amazing is that when creating these bots, instead of writing programming code, you have a conversation. That's right. You don't need to know about computers. So, if I want to create a book like this, and you try making it like this, then it's created in real-time. Of course, you need a developer's language to make the connections, but the basic things within GPT can be done naturally, as I just explained. It can be expressed in human words. I think I can explain that you can set characters, go beyond the limits of specific field information, and even connect to external services. So, I've looked into the reasons why it's good to write your own bot like this. Then, seeing is believing.

I don't think so. Shall we try making a chatbot? It will take me about 50 hours. 7 hours? Oh, that's great. Please share your thoughts on being able to delve deeply into attitudes and observations about the industry. After that, please summarize the related companies.

AI 트렌드 벌써 달라졌습니다...이걸 놓치면 안돼요 / 김덕진 소장 (1부)

"지식과 자산의 복리효과" 그 길을 복길이 여러분과 함께 걷습니다.
( 업로드 : 오전 11시45분, 오후 4시, 오후 7시 )
제휴 문의 : ystreet2021@gmail.com

인터넷 언론사 제호 : 와이스트릿 / 등록번호 : 서울 아53714 / 등록일자 : 2021.05.24
주소 : 서울시 영등포구 국제금융로8길 27-9 여의도 동북빌딩 910호
대표번호 : 02-6949-1701 / 팩스 : 02-6949-1702
이메일 : ystreet2021@gmail.com
발행인 : 이대호
Copyright © 와이스트릿  YSTREET ALL RIGHTS RESERVED. 와이스트릿의 모든 콘텐츠는 저작권법의 보호를 받으며, 무단 전재·복사·배포 등을 금지합니다.
