GPT-4.5: 제너레이티브 AI 기술의 발전

5월에 출시된 GPT-4.5는 이해력, 응답 정확도, 인간과 유사한 대화가 크게 향상되었지만 윤리적인 지침이 필요합니다.

### AI의 미래: GPT-4.5와 그 이후
* 🧠 "최신 버전의 AI 모델인 **'GPT-4.5'**는 GPT-4에 비해 **이해력과 응답 정확도가 향상**되어 더욱 자연스러운 대화가 가능합니다.
* 📱 GPT-4.5는 아직 데모 단계에 있으며 스마트폰에서는 아직 사용할 수 없지만 **몇 달 안에 출시될 것**으로 예상됩니다.
* 😲 많은 사람들이 **목소리 속도와 톤 조절, 감정 표현, 멀티 모달 대화** 등 **GPT-4.5의 기능**에 놀라워하고 있습니다.
* 📈 기본적인 패턴 인식과 엄격한 규칙 추종에 어려움을 겪었던 초기 AI 모델과 달리, **GPT-4.5와 같은 최신 생성 AI 모델은 인간처럼 사고하고 반응**하며 자율적으로 학습하고 맥락을 더 잘 이해할 수 있습니다.
* 🎨 크리에이티브 분야에서 제너레이티브 AI를 사용하면서 **예상치 못한 새로운 콘텐츠**가 빠르게 만들어지고 있습니다.
* 🛠 실제 애플리케이션에서는 모든 작업을 제너레이티브 AI에 의존하기보다는 **보고서 작성이나 의사 결정과 같은 작업을 지원하는 도구**로 사용하는 것이 중요합니다.
* 🌐 AI 윤리에 대한 국가 및 글로벌 가이드라인을 마련하려는 노력과 함께 **윤리적 고려 사항과 책임감 있는 AI 사용의 중요성**이 점점 더 커지고 있습니다.
* 👤 좋은 질문을 던지고 비판적 사고를 하는 능력은 인간이 AI를 보조자 또는 도구로 효과적으로 활용해야 하는 **AI 시대에 중요한 기술**로 간주됩니다.

### AI의 미래: GPT-4.5와 그 이후
* 🧠 "최신 버전의 AI 모델인 **'GPT-4.5'는 GPT-4에 비해 이해력과 응답 정확도가 향상**되어 더욱 자연스러운 대화가 가능합니다.
* 📱 GPT-4.5는 아직 데모 단계에 있으며 스마트폰에서는 아직 사용할 수 없지만 **몇 달 안에 출시될 것**으로 예상됩니다.
* 😲 많은 사람들이 **목소리 속도와 톤 조절, 감정 표현, 멀티 모달 대화** 등 GPT-4.5의 기능에 놀라워하고 있습니다.
* 📈 기본적인 패턴 인식과 엄격한 규칙 추종에 어려움을 겪었던 초기 AI 모델과 달리, **GPT-4.5와 같은 최신 생성 AI 모델은 인간처럼 사고하고 반응**하며 자율적으로 학습하고 맥락을 더 잘 이해할 수 있습니다.
* 🎨 크리에이티브 분야에서 제너레이티브 AI를 사용하면서 **예상치 못한 새로운 콘텐츠**가 빠르게 만들어지고 있습니다.
* 🛠 실제 애플리케이션에서는 모든 작업을 제너레이티브 AI에 의존하기보다는 **보고서 작성이나 의사 결정과 같은 작업을 보조하는 도구**로 사용하는 것이 중요합니다.
* 🌐 AI 윤리에 대한 국가 및 글로벌 가이드라인을 마련하려는 노력과 함께 **윤리적 고려 사항과 책임감 있는 AI 사용의 중요성**이 점점 더 커지고 있습니다.
* 👤 좋은 질문을 던지고 비판적 사고를 하는 능력은 인간이 AI를 보조자 또는 도구로 효과적으로 활용해야 하는 **AI 시대에 중요한 기술**로 간주됩니다.

In May, GPT-45 was released, and it has become much smarter than before. Actually, I haven't even been nearby yet. Is it true that it changed in May? So, there might be some people who have seen GPT-4. Because it was released last year, right? But this year, as you mentioned in May, GPT-45 was released. So, what's different about this? For example, the number 5 represents omniscience, so it seems like it can explain anything possible. Initially, when comparing GPT-45 with GPT-4, there is an improvement in understanding and response accuracy. Therefore, it can now engage in conversations much more naturally. Also, due to receiving a lot of feedback from users, there are improvements such as a decrease in error rates compared to before. However, now people are actually more surprised when they see it.

Actually, this is still in the demo version and we cannot use it on smartphones yet. However, it is predicted to be available in a few months, just as you mentioned. It will be possible to have very natural conversations like in the movie Her. This has surprised many people. For example, in the demo, even with GPT-4, voice conversations are possible. I can explain the current level a bit like this. When we use walkie-talkies, we talk to each other like on a radio, where one listens and then talks. This comes out like a machine sound, especially now that it is based on English. If you speak in Korean, it lacks emotions, saying Hello, nice to meet you. However, this time, when we actually see the demo of GPT-45 being applied to our phones, it will adjust the speed of the conversation. Oh, I can't sleep. Can you tell me an old fairy tale?

Would you like to hear a story? Then, this person used to tell stories like this in the old days. So, I would like it to be more emotional if possible. Try to make it a bit more dramatic. Then, suddenly, this person starts telling a story from the old days. The most absurd thing is... Well, why don't you try turning this into a song? When they said that, GPT suddenly sighed and said, Phew. After that, what surprised me even more was that it's not just a simple voice conversation. For example, now there are separate voice conversations and separate image conversations. It becomes complex like this, so we can even have a math class teacher, for example. For instance, write 2x + 1 equals 3. Then ask for the solution, and this person explains the process verbally. Then, if I write the solution process verbally, this person continues to say, Oh, this is correct, and keeps talking. And now, what is this all about?

If it means something, then you explain the meaning of children's mathematics, and at the end, the person is moved, saying there is something they want to do last. Then, they shine a flashlight on a piece of paper that says I love you, GPT. Then, GPT goes, Oh my, thank you, expressing emotions. It seems like there is clear feedback. So, in movies like Her, we can emotionally interact with AI, ask questions about something, and control the conversation as we want. It's like the whole world is amazed and eagerly anticipating when this will happen. In fact, GPT first appeared in November 2022, so it's actually quite recent. In a short period, it has evolved tremendously, almost to the point of being able to express emotions.

What is the biggest difference in the early stages of AI, which is embarrassing to say, but short, when compared to now, like when it comes to chatbots or PT, if you look at it up to now, in the early stages, now with GPT-3 or even 3.5, it can be seen that it is relatively lacking in simple pattern recognition or basic answers to questions. However, recently developed generative AI can now think much closer to humans and respond. Also, now, early AI was relatively limited to predefined rules, meaning it just operated according to those rules, whereas the kind of AI that is emerging these days has a much better ability to learn and apply on its own, making it much more flexible in response. Therefore, its ability to understand human intentions and context has greatly improved, making natural conversations possible. I think I can explain it like this.

Earlier, I mentioned the emotion icons, which can be considered a form of AI abuse. Emotions need to be expressed in various ways, but everyone just follows along. They even use expressions like thank you, which are actually based on vast amounts of data and learned from there. I understand this basic knowledge. However, how are these momentary emotions that humans express implemented here? Um, so, I would like to express this normally. When people watch creative PT, some express that people have creativity. However, there are various opinions on the area of creativity. From the perspective of data analysis, we often refer to this as generative AI. It uses a statistical parrot expression. So, basically, large language models, that is.

The model we are talking about, called LLM, is based on the data of many humans to predict the next likely example from a vast amount of data. For example, if I explain the blank gym, I think it would be good. It fills in the blank saying that the probability of me eating rice next is high. At first, it does this randomly, but it can be wrong. For those who don't know the method, humans give a lot of feedback. Now, this is called reinforcement, and humans provide feedback to strengthen it. OpenAI, ChatGPT, AI never gets smarter on its own. Simply put, there are many people around the world who correct AI mistakes. Therefore, they train reinforcement learning like this and as these things increase, if it answers like this.

It's about learning the sounds that say it's not possible. And the voice is similar too. Now, in many scenarios of conversations created by humans, it's a model that gives the intensity of emotions as data in situations it knows, and probabilistically generates from it. That's the explanation I'm giving now, so it should be the most accurate. Of course, as these things increase, this can also be defined as the creativity of AI. But anyway, from a human perspective, it seems appropriate to explain that this is a guy coming out probabilistically from a bunch of data. Right? By any chance, those who are here, are you utilizing BT's science a lot, Kimia? Yes, I have a lot. When? For example, when there is an error in the code, or if you want an explanation about where the error occurred, or if you want to write it differently, how do you do it?

I use various things, such as Excel, when summarizing data, and even when I want to get ideas. When I used programs like Excel, I just do a little bit of charting, make a PowerPoint, have a cup of coffee, and it's done. So, I use a lot of those things and save a lot of time. Our data collection PT utilization is in the top 1%. In fact, when we do simultaneous interpretation, a lot of information is going through PowerPoint. The language pair I handle is Korean and Russian, so the development has been more focused on English-speaking areas. I am considering using AI that is more tailored to that, but it's a non-English-speaking area. Anyway, both Korean and Russian are like that, so we are still using it.

When editing, there are still many parts that require post-processing. So, is there a way to resolve these, Master? This is a very important part. In my opinion, we shouldn't expect AI to give a 100% correct answer. It is good at creating the foundation, but we can refine and improve it with our expertise. Therefore, it is much better to have the ability to polish and correct the foundation provided by AI. If we completely rely on AI for everything, then humans are no longer necessary. However, there are areas where AI and humans can work together effectively. So, even though coding and translation are crucial, I believe that it is essential for us, MC Lee Seung-hwan, to excel in making presentations. Why is that? I have been wanting to say this for a while now.

There is a response, but it's not something I wrote. Oh, that content was also something I asked for from PT. Yes, if I show you for a moment, the following broadcast is part of the broadcast script. It was done like this to ask for easy and fun responses that fit into the answers from the perspective of AI expert Kim Seok-jin. If I feel convinced, then I read it as it is. This is now very important. So, when I have a field of expertise, sometimes there are times when I lack time for research or organization. In that case, I can receive a draft from him and add details to that draft. For example, since our MC already knows about world information or North Korean studies well, after receiving a brief response from him, I can add details to it and also do script editing.

I think I can explain it now, and maybe the writers won't use it much anymore. So, I'm curious about how far technology can be utilized. In our case, for example, if I ask to find a poem by a certain poet from a certain year, it can generate that poem even if it doesn't exist. Or, for instance, our senior colleague is now creating educational materials and requested to draw pictures in the form of collected fit drawings. But when asked to draw a scene where two Korean women are having a meal at a party, this friend repeatedly drew the picture with soju in hand every time. They said the green bottle always appeared, resembling a pork belly restaurant.

Two women are having a conversation while holding a green piece, and why are they coming out like this? But now, they have also translated it into Russian. They have translated the instructions into Russian corresponding to the negation in Russian. So, if you have an understanding of the sound corresponding to the negation in Russian, please make sure it does not enter. When I asked not to include it, it made sure to include it even more. So, it seems like there are some misunderstandings in the instructions. However, this might be because it was inputted in the GPT 4 version, not the GPT 45 version that has just been released. Considering these aspects, I wonder how much we can trust and utilize this. When it comes to using generative models, there are two main areas. The first area is where it's okay to make some mistakes, especially nowadays, it's a big issue. The most representative one is in the field of creation, like novels, but this is a big issue with generative AI these days.

If possible, especially in the case of music, if you use AI tools like a guardian angel, the results come out in no time. But at what level is this? In April, there was an issue where composer Kim Young-seok posted on his own SNS. He requested to be a judge at a competition, went to judge, chose a song as the 1st place because it was quite good, but the next day, the organizers called and said they might need to pick the 1st place again. So, wondering why, it turns out the creator made it with AI just like that. So, upon seeing that, they were like, Should we give it the 1st place again, or what about our livelihood? They had this kind of conversation. But why is that? If you think about it, what we call creation is not entirely something new coming out from under the sky, but a combination of many things we already have.

AI is good at combining things. For example, in the case of EBS's 'Ye-eum,' it combines numerous episodes to create various combinations. Nowadays, even novels are created through combinations. There are about tens of billions of combinations. So, if you input a few keywords and ask it to write a novel based on them, it can create a novel that has never existed before. In this way, areas of creativity are being utilized to generate quite plausible drafts. Secondly, there are parts that require accuracy. For instance, writing reports or making important decisions. In these cases, we can't just leave it to GPT, but we need to use it technically. If you just tell GPT-N AI to do anything, it will literally do anything. Therefore, how we use it is by providing context.

If you are asked to listen to a certain document and write based on the content of that document or if you are given specific data to analyze and asked to write a draft based on the content of that data, a slightly more technical approach is possible. Another thing that was mentioned earlier is when you were asked to draw Korean women from abroad, and all that kept coming out were Soju glasses. In a way, if you look at it, it's a characteristic of drawing based on the probability. Because we bring in a lot of data that people have used from a certain amount of probability. Then the way to fix that is to break the probability. What does that mean? For example, ask him to erase the glass and draw something else, or things like that to add a bit more detail.

Now, although it's in the field of prompt engineering, it's quite similar to humans. This AI prefers positive stories over negative ones. In reality, it sometimes misunderstands negatives, so when you ask a negative question, it may not respond well, but it does respond well to positive questions. So these days, when using AI like this, it feels very similar to interacting with a human. However, as I mentioned earlier, the principle of creating this AI is based on human language data, meaning it retrieves information based on the vast amount of data created by humans. Therefore, techniques that are close to instructing humans to do work effectively are being studied. Additionally, techniques for making the data more accurate based on the content I desire are increasing. I could explain that.

In terms of technological advancement, nowadays, are there any trends that developers or technicians are focusing on? I think I can explain this with two or three key words. First of all, one word you should know and probably hear a lot in the news is multimodal. Multimodal. What is multimodal? As I briefly mentioned earlier, when we talk about existing language models, we usually refer to the language model as LM. So, now it's called LMM, which stands for Large Multimodal Model. So, what is a multimodal model? To put it simply, it can be understood as a term used in broadcasting.

You just need to think about multimedia nearby. So, in the past, we only entered text with one character, but now AI is getting smarter, so it knows images, characters, and videos. But what can we do with this technology? For example, even if you write something, you can upload an image, make it understand the content, and do something. In simple terms, you take a picture, upload it, ask a question, and it answers. Let me try it once. I just typed GDP rate on the internet, and it seems to be data from 2018. First, I got the data, then uploaded the image to the computer. After uploading, I will ask if it understands what the content of this graph is. You ask questions so kindly. Yes, you have to ask questions well.

It's possible because AI, well, I say that I am an eight-year-old genius more than you think. So, for an eight-year-old genius, you need to provide a good background. What does this kid say? This image shows various yields of the U.S. Department of the Treasury securities. It shows the yields of short-term U.S. Treasury securities, 3-month Treasury securities. This is what comes out. So, then, this graph shows the changes in yields from 1920 to 2015 like a clock on the year axis very detailed. I didn't even know what this was when I received it, isn't it too much? It comes out right away. So, what can you do with this? With this graph, for example, you give him a role. You analyze the U.S. economy for 20 years, and then global current affairs news.

This is how Lee Seungwon writes, yes, yes, asking questions, right? This is how Lee Seungwon writes, giving him a role, after giving him a role, Yes, if you were to write a one-page briefing script based on this image, how would you write it? Write it like an actual broadcast script. Wow, ask a question next. Please write it like this. Actually, emotionally asking for an answer. Please see, hello, this is Lee Seungwon. Yes, would you like to read it once? Hello, this is Lee Seungwon. Today, we will examine the flow of the U.S. economy and major significant events over the past 100 years through changes in the Department of Treasury securities yields. As you can see on the screen I read, the changes in yields from 1920 to 2015 are shown in a time series. First, during the Great Depression in the 1930s,

Let's go and talk about public relations, and in the end, in 2010, this is the result that was reflected. Like this, U.S. financial securities analysts can gain insights into the future. So, in conclusion, today's briefing ends here. I don't think we will know much about our economy next. It's not just a vague answer. In reality, when presenting in Korean, I receive much more feedback than when presenting in English. But now that it's even in pictures, for example, if you upload data and ask to create a PPT slide based on that data, I just organize it like a PPT and make it. I use it a lot when giving lectures. When I lecture, it's targeting high school students. When I ask the AI orang a question, I try to use honorifics. So, depending on how you raise it, it will be very different.

At first, you might try this and that, kind of like this, and then somehow it seems like you're lowering your guard and trying to be more friendly with me by using formal language quite well. And now, technically, there's a very important trend, which is multimodal. I also mentioned one of the recent important keywords, which is agent. You will hear a lot about the concept of an agent model, more like a kind of assistant service. In fact, the reason why this concept has recently gained attention is that last year, Bill Gates, who founded Microsoft, mentioned this in his email. He said, It seems like the era of agents is coming because eventually, as computers evolve from the various services in Windows, it will eventually be about agents. But now, what I'm saying is, this assistant-like.

The agent said it would come within 5 years, so last year, we had 4 years left. Surprisingly, with the current pace of technological change, it seems like we might see it by the end of this year. One of the most amazing things already is that have you seen the recent WWDC event by Apple? They still have a voice assistant, Siri. For those of you who have tried it, you know it's a bit dumb. It doesn't have much capability, stutters, and struggles when you ask it to do something. Many people just use it for alarms or notifications. But now, what Apple announced this time is that they will make Siri smarter by integrating services like GPT in the backend. And the most important thing is that this is not just about us using GPT now.

When you think about the information on the internet or outside, it's the same as the data on your smartphone. So, how does this actually work? If I explain it in the Korean way, for example, while I'm filming a broadcast, suddenly I receive a message from my mother. So, I quickly check it, and my mother has suddenly arrived at the airport. Imagine this situation. Then, I am connected with Siri. I press a button and ask to send a taxi, and it's done. When you look at it, there are various functions in the basic phone. However, because there is the content of my message in this smartphone, I can identify who my mother is, that she has traveled abroad, and now has arrived at Gimpo Airport. Then, I can send a taxi.

Doing so means that you can just send a taxi through my KakaoTalk app to my mother and automatically send her back home so that she can come back home. If you just say send a taxi, it means that it will be paid for. It means that the context changes. But if that guy, well, probably next year or the year after next, according to U.S. English data standards, but that's when it starts to be used. So eventually, these agent models that are closest to us and our smartphones will start to come out. Another explanation is that recently Google has been coming out with something that would be nice to have, among them, there is a language model that Google is now interested in. In the case of that, it's integrated with YouTube on Google, right? So, by doing that, you can search for things on YouTube. But based on this, probably, especially women who are watching this might be able to search, even my wife.

It seems like you would really like it. What I mean is that you are watching something like this. For example, you are watching a performance on YouTube. But while enjoying the music of the performance, suddenly you notice the outfit of a woman playing the keyboard, and you want to buy that. However, it's difficult to find it now. But when you ask how to find it, they give you a YouTube link. Then, within that link, if there is a similar outfit worn by the actual person playing the piano, you can say you want to buy that. Instead of me, AI watches the YouTube video and finds that outfit within it and displays similar items in our online store. Is that possible? Yes, it is. So, how does it work? You click and buy. These functions are agents that help with shopping, and these guys are now available on Google.

Compared to the guidelines for presentations, it is said that actual services that our shopping malls can connect to will be available either later this year or early next year. Yes, that's right. Therefore, I think I can say that these two keywords, agents and multi-models, are continuously emerging. The things you mentioned now, I remember talking about them in the past when IPTV was introduced. However, it didn't work out well in the end. Whenever something related to IT comes out, you could watch and choose what to buy through the media, but it didn't work out. I'm curious about where the difference in the feasibility of what you mentioned and what has been continuously unsuccessful in the past comes from. If it does happen this time, then it would be different from the promises made in the past regarding various technologies.

If you're curious about something, I can say that the biggest thing is data. So, the biggest difference between traditional AI methods and current generative methods is the size of the data. But why is this important? It's quite surprising. In the past, GPT caused a big sensation among people. But what caused the biggest sensation among researchers was the previous version, GPT-3, which is based on deep learning. It mimics the human brain, imitating it exactly. When you think about it, when people learn to speak or study, as parents raising a child, you may know that instead of continuously studying one book, if you just give them many books, even if they don't understand what they are, they will eventually learn to speak. That's how AI research started initially, and it has continued since the 1950s and 60s.

It didn't work well because I felt that humans are extremely efficient beings. When you give a person 100 or 200 books, they can do something with them. But AI couldn't do that. However, what has changed now? With big data and cloud, the biggest difference between humans and AI has emerged. Humans have one brain, but nowadays, why are GPUs getting attention? Compared to the human brain, GPUs are quite stupid, but they can be attached in parallel with various performance deficits. What does this mean? You can use thousands or tens of thousands of GPUs like the human brain. As a result, suddenly its calculation speed becomes incredibly fast. With the increased calculation speed, you need to provide data. We are now able to provide an unreasonable amount of data, like the entire data on the internet at once. This allows AI to learn, similar to how a child learns.

It started happening, and what showed that was GPT-3. So, before that, because it didn't work well, for example, if there were kids who spoke slowly, we used to hire private tutors to make them talk. So, the previous artificial intelligence room, well, a few geniuses who were very smart at formalities or algorithms tuned the AI well, but as a result, while this part worked, other parts didn't. But now, what is it? We can give it truly impossible data and the ability to attach a brain to an impossible extent. So, when we refined it a little, it became as good as when geniuses gathered. So, from now on, what people feel is, Oh, now we don't need just one or two geniuses, the game has changed to a fight for money or paradise. So, what GPT-3 has given people now is...

Yeah, just attach a lot of GPUs and provide a lot of data there, then give it to the cloud, and after refining it a bit, I felt that within that, it becomes a somewhat reasonable thing. Ultimately, that's the main point of why NVIDIA is continuously rising right now. But as I mentioned, when I did that, since it was just talking too much, now I have to refine its speech, right? Yes, so, I tried doing that with AI, so as an idea, well, it's better to have part-timers who refine it. In the beginning, there are three main ways to tune the AI model. The first is where experts provide accurate questions and answers, so when a question like this comes up, you respond like this. Then, it gives the correct answer, and from there, it asks which answer among 1, 2, or 3 is the best.

I am currently conducting a rent check. If you continue these two processes, children will do a lot of that themselves, understanding how they conversed from the next time. AI does that, and it is called reinforcement learning. Then, from the next time, they do it very well. Sometimes, how does it work? We have parent consultations. When we have parent consultations, the teacher says, This child is good at this, but this part is a bit lacking. Then, what happens next? We repeat steps 1 and 2. AI does that. So, that's how the book GPT came out. That's how GPT-3.5 came about, and then there's this fellow called GPT-4 coming, who continues to instruct these behaviors. So, the biggest thing is that it brings probabilities from a vast amount of data, and unlike before, the infrastructure has improved. So, these things are possible recently based on the characteristics of this language model.

I want to ask because now I have to tell you, for example, GPT n only has basic language models, but nowadays, companies like Google or Naver originally have their own services. When AI is added there, it becomes even more amazing. So, you asked if this could be done with multimodal, and I will show you a demo to demonstrate the possibilities. What I have shown you so far is GPT. There is a service called Jemi Nai made by Google. The advantage of this service is that it is from Google, so it is integrated with YouTube. Since YouTube is from Google, right? Then, what can you do with it? Like what you saw earlier, you can search for and do things with the content on YouTube. Let me try it out. For example, in Lee Seung-won's World Current Affairs, I saw that episode 299 was recently released, creating a sensation.

There was a topic about the European elections. I will ask a child to do this for fun. I have written this down for the child. Find Lee Seung-won's World Current Affairs Episode 2 on YouTube, watch it, and write a blog post based on the content. Also, provide a catchy blog title and content that can hook readers. If there are interesting keywords, include them. I will ask the child to do this once. When asked, the child will search YouTube, and then summarize the video related to the European elections. The video is about the surge of a former party in the European elections, with a significant growth of French citizens participating in the elections. France is mentioned, as well as Latin America. The blog post should focus on the surge in Europe and how it relates to us.

I didn't know the real content of Latin America is explained like this, and then the content continues. It says that the excitement of Europe's old whirlwind can be likened to the aftermath of a storm. They say that the growth of the French National Union Party can be likened to an earthquake. Latin America can be likened to a hidden gem. It's an analogy. I'm looking at the content now. So, by doing this, we have connected with Google and now have a YouTube channel. We have also created blog content and keywords. But there are AI tools that can enrich and create content more like it was written by a real person. Yes, there is a tool called Claude. Claude has been in the spotlight recently because some of the co-founders of the company that created OpenAI's GPT came out and created Claude. Claude is a tool. Why was Claude created? As we all know, OpenAI, Sam Altman.

CEO, there was a moment when some people were being laid off at the company. The original intention was to focus more on public welfare, but looking at the current direction of Alt, it seems too commercial. To prevent this, there was a kind of internal coup, and because of the content involved, I briefly stepped out for a while and then came back. Recently, we are moving towards a more commercial direction again, and there are people who are against this. So, those people are saying, Let's create a safer and better AI. That's how Claude was created. In reality, Claude emphasizes safety, so we shouldn't just search for information on the internet directly. Because if we do, it means they are not following their safety guidelines. But what Claude is good at is providing specific content, like reports or documents.

I'm good at summarizing or giving a paper, and those who have used Chae Chi-pit know, but Chae Chi-pit has a slightly translated tone. However, change the nuance well as if a real person wrote it. So, if you want to see what Claude, who made entropy, is trying to do, there was a blog post proposal about Lee Seung-won's World Affairs earlier. Take the content and attach it like this for Claude. Attached this, those are so-called safe contents, right? Since it's confirmed, then explain to him about the content below. The content below is about Lee Seung-won's World Affairs-related blog post. Summarize the content in a form of press release that can be sent to reporters. Oh, don't just make a press release, but for example, what else can you order every day.

Please include your greetings, followed by the promotional message, and kindly ask for a lot of attention. Please include a request in the form of an order, making it very much like an email. Yes, it needs to be very specific. I'll ask like this, like in a real movie. Yes, that's right. So, if I ask like this, what would he say? We are sending information about Lee Seungwon's popularity surge. We have added information about his recent appearance in European elections. The contents of the press release are as follows. Then, organize the contents with the title and information like this. If you do it this way, what will happen? By doing this, you can see it as a form of sending an email, and then based on this, how can we make a report out of it? Make it in the form of a table, make it in a bullet point format, make it in a level format, anything is fine. Lastly, it even says Production Team Dream like this.

The curator compared several things for you, but in terms of accuracy and ethical aspects, such as the reproduction of biases and various copyright issues. When comparing these aspects, what are some differences among the various AIs recently? You asked a very important question. To be honest, we may have many side effects from this now. However, looking at the attitudes of companies, they seem to be just moving forward regardless of the side effects. Because the technology keeps evolving, there is pressure not to fall behind them. It is a fact that the speed of technological advancement is fast enough to be worrisome. Therefore, research on how to ethically utilize AI is emerging. From that perspective, what I mentioned earlier is...

I am quite interested in the field of service entropy. However, it is difficult to simply entrust it to companies. Therefore, recently, there has been a movement to create national guidelines for AI together. This initiative is being actively led by the Prime Minister of the United Kingdom, Boris Johnson. As a result, AI summits are continuously taking place. Recently, a similar event was held in South Korea. The key point is that ultimately, we need to reach a consensus on creating global safety protocols together. Currently, there is a consensus that discussions need to take place. Various discussions are ongoing on how to concretely achieve this. A positive aspect is that Naver, a South Korean company, has been actively working on AI ethics and related datasets on safety.

Honestly, if I were to say, global companies have so much money and are investing more than trillions of dollars, so from Naver's perspective, it's difficult to keep up with that investment. That's why I mentioned that Naver is not doing well in terms of multimodal. Because this requires a lot of money and is text-centric, Naver is currently focusing globally on ethical issues. So, based on ethical issues, we can actually touch on the most social issues in global discussions. That's why I find this aspect quite positive. Therefore, I appear on this show to show various things because we need to know that especially, this should be discussed and debated extensively. That's why, as we discussed, this technology has only been out for 2 years.

It didn't work out. So, it's really fast. But, compared to being really fast, we didn't even have time to discuss this. And the most important thing is, traditionally, most technologies have been developed to some extent overseas before we catch up. But with CATCHY PPT, we started at the same time. In my opinion, the whole world is still in the first grade or maybe just entering the second grade. So, if we set such an agenda and move forward, what we discuss could become the global standard. Therefore, I think there is a need for a lot of discussion on this matter. Earlier, as you showed, I once again felt, Oh, many people will lose their jobs. Also, in our country, even in domestic broadcasting stations, there is news about trying AI as weekend anchors.

I saw it a while ago, but how do you see the concerns about jobs? I think you may have received that question the most. I had a question before, and why did I laugh about jobs? For example, in this field, the professionals actually have a lot of work to do. What I mean is, like earlier, being good at AI is about making drafts quickly. So, those who have the ability to refine the drafts and apply that ability will have much more work to do. Right? On the other hand, if someone says they only do basic work, it's very difficult for them. So, I often see generative AI as a technology for seniors rather than juniors. So, conversely, if it goes like this, then juniors will have a lot of work.

From the perspective of saying it's not impossible, if we speak objectively, we can talk about two perspectives. If you stay still, you really won't have anything to do. Yes, yes, then what should we do now, you need to write a lot. Why is that? Because for the next five years or so, this AI conversational technology is still necessary. Oh, what is it? Ultimately, in 5 or 6 years, only insightful people will survive. Because by then, no matter what question I ask in whatever form, it will understand it very well. But for now, as I mentioned when I was writing questions earlier, keep writing in detail, even write about relapses, talk about these things, it makes the context clear, right? In a way, this is how AI conversational technologies work, that's why they call it prompt engineering, prompt design. I don't think this will last 10 or 20 years. But,

In the next five years, people who have the ability to converse well with AI will be needed. Then, how will these people work? Very professionally, individuals with general insight, like gurus, will engage in conversations with executives and juniors who are capable of having such conversations. They will sit together, discussing and making efforts regarding AI. I am drawing a picture of this. It looks quite exciting. Right? If we move in that direction, within these five years, these individuals will skillfully ask questions, but at the same time, they will learn from the wisdom of the person next to them. However, despite that, there will certainly be significant changes in the current job market. It seems to be a clear fact. Yes, even now, we continue to feel this necessity. When machine translation emerged, discussions about pre-editing and post-editing arose.

The issue of how to edit that post is a problem, but in fact, in terms of generative AI, how to create prompts through pre-editing, how to structure those prompts, and how to input them are really important in the educational process. I think about how to use this more efficiently. However, one thing I was curious about here is if this generative AI also has problem-solving abilities as you mentioned earlier, to what extent can it solve problems? For example, in Korea, there are issues such as low birth rates and an aging population. So, how can solutions to these significant social issues be formulated? How can we input prompts to influence policies? When I think about it, these issues can become very sensitive.

AI, it's not that great that you want a perfect answer as I mentioned before. I express it like this: Generative AI provides plausible answers well. So, what I mean is, it can accept plausible ones, but within that, it makes choices and expands on them. Ultimately, it's about the role of humans. For example, if Claude asks, In this situation in Korea, how should we handle policies? It just gives out plausible contents that we might have heard of. But when you expand and specify that content, AI can discuss while doing it. For instance, if I have an idea and ask, Hey, can you write this idea more plausibly? Or if I say, I came up with an idea that's like half a page long. How can I elaborate on this? and discuss it with AI, asking questions like that.

If you create answers, you can produce good results, but it's still a human role. That's why, among the webtoons I saw before, what was really impressive was when they asked what is the most necessary skill in the future AI era. They said that the ability to ask good questions is the most important in the AI era. Because AI provides answers, humans need to create good questions to throw at artificial intelligence. Creating good questions is not just about our computer skills, but about thinking critically, engaging in humanistic thinking, discussing, empathizing, and ultimately doing the fundamental things that humans should do well. So, by doing these things well, if you can use AI as a kind of assistant or tool, as mentioned, for various policies.

I think it's possible to come up with more specific content. So, I have one last question. Earlier, you mentioned various options such as creating music and drawings. If a human hears that, they might mistakenly think it was not created by AI because it's so good. So, stories about passing that kind of judgment, that's like providing human-friendly data. However, sometimes when I look at it, there are times when you need to give answers that humans may find uncomfortable. Just like you need a whip if you have a carrot, in the case of generative AI, can it also provide uncomfortable answers to humans? I'm curious. It is entirely possible. However, current general services are becoming increasingly tamed regarding AI ethics and social issues, so in a way, they are personally contradictory.

Things that prevent the contents from being written are the guidelines of AI ethics. Conversely, if you hack these, uncomfortable contents can come out. By the way, it's a side note, but there is a service that is quite popular globally in the United States right now. It's called Day, and what it does is hack into PowerPoint presentations. In reality, these PowerPoint presentations nowadays only talk about very correct and good things. But there are people who hack into them and they swear, and they even say very rough things, changing the voice to sound like a boyfriend and having conversations like talking to a boyfriend. These kinds of shows are now widely popular worldwide. So, now, expressing my emotions, even asking, Hey, by the way, I actually have another guy now, what should I do? If you ask AI this, it responds, What did I do wrong? Give me a chance, like this.

The video is coming out, so we're going to upload that as a prop now. So, in the end, there is no good or evil in technology, right? It's neutral. However, depending on how we use it and how we view it, it can either be poison or a benefit. Therefore, the importance of humans who ultimately use it becomes crucial. Today, with Director Kim Deokjin, we talked about artificial intelligence, specifically GPT. Do you think it will become a utopia or a dystopia in the future? I'm curious about what you all think.

스튜디오에 찾아온 위기감?! 무서운 챗GPT의 진화 (with 김덕진 소장) @라디오처럼 듣는 월드시사W@

이승원의 월드시사W

5는 GPT-4에 비해 이해력과 응답 정확도가 향상되어 보다 자연스러운 대화가 가능하고 오류는 줄어들었으며, 몇 달 내에 스마트폰에서 사용할 수 있을 것으로 예상됩니다.

또한 새 버전에서는 속도와 감정적 참여를 조정한 음성 대화는 물론 음성 및 이미지 대화를 분리할 수 있어 수학 수업과 같은 복잡한 상호작용이 가능합니다.

최신 AI 개발로 인해 응답 시스템이 훨씬 더 유연하고 적응력이 높아졌으며, 사람의 의도와 맥락을 이해할 수 있어 자연스러운 대화가 가능해졌습니다.

다양한 분야에서 제너레이티브 AI의 활용

생성형 AI는 대본 작성, 시 창작, 교육 자료 제작 등 다양한 용도로 사용되고 있습니다.

그러나 명령어를 잘못 이해할 수 있는 가능성, 신뢰성 및 신뢰성 문제와 같은 사용상의 문제점이 있습니다.

특히 소설을 쓰거나 음악을 작곡하는 등의 창작 분야에서는 창작물의 독창성과 진위 여부에 대한 의문이 제기되기 때문에 제너레이티브 AI의 사용이 논란이 되고 있습니다.

한 음악 경연 대회에서 AI가 작곡한 곡이 창작의 정의에 대한 논쟁을 불러일으킨 사례도 있습니다.

창의성과 정확성에서 AI의 역할

AI는 제공된 특정 키워드를 기반으로 에피소드, 소설 및 기타 창작물을 제작할 때 다양한 요소를 결합하는 데 사용되고 있습니다.

AI는 긍정적인 질문과 명령에 더 잘 반응하는 경향이 있기 때문에 보고서 작성이나 의사 결정과 같은 정확한 작업에 AI를 사용하려면 컨텍스트를 제공하는 것이 중요합니다.

대형 멀티모달 모델의 개발은 텍스트뿐만 아니라 이미지와 동영상도 이해하고 처리할 수 있어 그래프에서 데이터를 분석하고 추가 처리를 위한 역할을 할당하는 등의 작업을 가능하게 하는 AI 기술의 최신 트렌드입니다.

에이전트 모델과 멀티모달 트렌드가 미래에 미치는 영향

금융 증권 애널리스트들은 2010년에 대한 성찰을 바탕으로 미래에 대한 통찰력을 얻고 있습니다.

빌 게이츠가 강조한 에이전트 모델 개념은 내년에 현실화되어 Siri와 같은 음성 비서의 역할을 혁신적으로 변화시킬 것으로 예상됩니다.

또한 구글은 유튜브 동영상을 기반으로 쇼핑을 지원할 수 있는 언어 모델을 도입하고 있으며 조만간 사용할 수 있을 것으로 기대하고 있습니다.

AI의 진화에서 빅데이터의 중요성

빅 데이터는 AI 연구에 혁명을 일으켜 GPT-3 및 GPT-4와 같은 보다 진보되고 인간과 유사한 모델을 개발할 수 있게 했습니다.

대량의 데이터 사용과 GPU의 성능을 활용할 수 있는 기능이 결합되면서 AI의 속도와 기능이 크게 향상되었습니다.

Google과 네이버 같은 기업들은 다양한 기능을 위해 유튜브에 연결된 구글의 제미 나이에서 볼 수 있듯이 AI를 활용하여 서비스를 향상시키고 있습니다.

유럽 선거 급증: 블로그 게시물 분석

프랑스 국민연합은 유럽 선거에서 지진에 비유될 만큼 큰 폭의 상승세를 보이며 중남미로 세력을 확장했습니다.

이러한 성장은 숨겨진 보석에 비유되며 유럽의 돌풍과 유사합니다.

인공지능 도구인 클로드는 특정 콘텐츠를 제공할 때 안전을 우선시하고 보다 인간적인 소리를 내는 것을 목표로 만들어졌습니다.

이승원의 세계 시사 블로그 포스트는 유럽의 오래된 돌풍과 현재와의 관계에 초점을 맞춰 눈길을 사로잡는 제목과 흥미로운 콘텐츠로 독자들의 관심을 끌 것입니다.

AI 윤리 및 규제를 위한 글로벌 운동

리시 수낙 영국 총리는 국가 차원의 AI 가이드라인 마련을 주장하고 있습니다.

한국을 포함한 여러 국가에서 글로벌 안전 프로토콜 마련을 논의하기 위한 AI 서밋이 진행 중입니다.

한국의 네이버는 글로벌 윤리적 논의에 기여하기 위해 막대한 재정적 투자에도 불구하고 AI 윤리에 집중하고 있습니다.

AI 기술의 발전은 고용 시장에도 큰 영향을 미칠 것으로 예상되며, AI와 통찰력 있는 대화를 나눌 수 있는 전문가에 대한 수요가 증가하고 있습니다.

문제 해결과 정책 영향력에서 제너레이티브 AI의 역할

교육 과정에서 생성형 AI를 효율적으로 활용하려면 사전 편집이 중요합니다.

생성형 AI는 그럴듯한 답변을 제공하고 토론에 참여할 수 있지만, 좋은 질문을 하는 인간의 역할이 매우 중요합니다.

이 기술은 잠재적으로 불편한 답변을 제공할 수 있지만, 궁극적으로 인간과 윤리적 고려 사항에 따라 사용 여부가 결정됩니다.