Today, I'm not here to talk about a specific definition, but I would like to talk about some trends and tendencies, not only in Korea but also globally. With the emergence of chatbots and GPT-based AI, the biggest trend is in search, where search and knowledge are being heavily utilized. Many people are still benefiting from these search-based AI models. This may seem obvious, but it's not just about search and chatbots when it comes to these AI models. From the beginning, I have mentioned that there is more hidden behind these AI models than just search and conversation. So, besides the knowledge that search and chatbot-based AI possess, there are many other contents hidden behind these AI models. Therefore, a representative example is how I create PowerPoint slides, how...

If you ask it to create a PowerPoint based on a document, it will create a PowerPoint using the document as a reference. It takes about 1 minute and 40 seconds. Of course, this time is based on the current standard, so it may get faster or slower, but can anyone make 18 slides of PowerPoint in just 1 minute and 40 seconds? It's not easy. Even if the PowerPoint is not really great, excellent, or perfect, it should have a certain level of quality that can be used for discussion. I think it's okay to spend 1 minute and 40 seconds making a PPT. We will make the PowerPoint based on a file. So, what is this file? Oh, it's about the global carbonated beverage trends in 2024. We will make a PowerPoint about that. Maybe someone presented it at a company.

I roughly predicted that I would go to the company that makes cider and explain, but it seems like the video is too long to do a demo like a company, so I didn't cut or edit it in the middle. I recorded it as it is. I will make it with this kind of table of contents and show it without being boring in the middle. However, most of the employees of our company also saw it, so even though I said to make it, they are not looking at it like you guys. They put down PowerPoint and do other things. They multitask. There is no one who is watching it interestingly like this anymore. It really becomes multitasking. Now, I'm going to make a slide note. I made 18 PowerPoint slides in 1 minute and 40 seconds. The introduction image is not even something I put in.

In just 1 minute and 40 seconds, I will show you how to create a Word document like this. How is it? Actually, it's amazing. Because I am not a human, without refreshing drinks, the fact that I can create a document like this alone is surprising. Also, it seems like this type of document creation could be most useful for office workers. Furthermore, for those creating various creative content, even in the field of making videos, many have seen how OpenAI's GPT-3 project can produce amazing videos when prompted to create a stylish woman walking the streets of Tokyo. Moreover, not only recreating parts of existing videos, but also creating within this video.

Special information, which includes spatial data, is also included. So, this woman creates even those physical parts when she shakes her bag, the bag naturally shakes. Of course, there is a limit that only about 30 seconds of video can be produced so far, and there are some issues with various physical aspects. However, many people were impressed. Isn't it too natural? With a bit of slow motion, the angle at which the bag is held, and the appearance of the light projected on the water, but if you look closely, the letters come out in a slightly distorted form, not the usual circular shape. However, there are some disadvantages. Nevertheless, many people were shocked. Someone asked to create a similar video and the time it takes to do that.

How much money are you going to talk about? But you guys, please give me a week, and I need about 30 million won. I talked about airplane tickets, hotel expenses, shooting fees, editing costs, and so on. But suddenly, someone said, If you do it in 30 minutes, just give me 20,000 won. The most important thing is time. It may take a week to create the video from the conventional perspective. If it can be done in 30 minutes or 2 30 minutes, the conversation changes. Also, from the perspective of someone making a program, if you see it like this, Please write the multiplication table code, even if you are not a developer and have created some code, now, wait a minute, if you don't understand what someone is saying, you need comments to provide this to someone. I asked to add comments, and then from 1 to the end, repeat.

Output the results of the repeated request. This is the multiplication table, right? Not only that, but also create test code for this to practice. What this means is that the thing developers hate the most is that even though the developers have written all the code, they have to write all the comments. They get frustrated having to write each one. But to check if this works properly, you need to write unit test code. So, developers end up working overtime sitting in front of the computer all the time. This is not a great thing, right? The price of this agent, although I didn't come to sell products, I think I should mention the price. It's $19 US dollars. So, if you use it for a month, it's $19. In Korean currency, it's about 30,000 won. So, ask the people around you who develop programs, tell them, I'll give you 30,000 won a month to write comments on the code, create test code, and write documents.

So what happens? Relationships are severed, contacts disappear, history is caught, right? I am doing boring and uninteresting tasks. Will people be involved or will people do creative work and leave those mechanical tasks to AI? In fact, over 60,000 companies worldwide are using this function. When was this service released? It was launched in November last year. It was launched last November, and after that, it was released, but already over 60,000 global Fortune 500 companies are using it. However, now many domestic companies are also using it. There may be some of you using it here. But in Korea, it costs $19 per month. If Microsoft does AI, most people here say, In the end, they have shown their true colors. So what is the value of this light? So, in the beginning, we started with this generative model.

When it first came out, it was all about the basic model, like GPT 3.5, Turbo GPT4, and so on, talking a lot about basic models. Now, more extended specific domain models are starting to emerge. For example, in the medical field, there are models specialized in medicine, and in the legal field, there are models specialized in law. Also, the content format is not just text anymore. As I showed you a little while ago, not only code but also videos are starting to appear, and from the early days of sound, there were AIs creating music. Now, there are AIs that can create 3D models. They can design exhaust pipes in 3D, adjust shapes and angles a bit more, and these things are coming out now. So, what's good about this? It means that not only professionals but also non-experts can create 3D objects. Of course, more specialized knowledge may be required, but that's why.

Such parts are now not only creating data and creating content but also transforming the organization accordingly and generating various data tailored to the characteristics. Next, when viewed from a business activity perspective, internal communication within companies is changing, followed by changes in collaboration methods, and then the systems managing the entire content are changing. It's not just about looking at this based on AI, but based on this, generating various forms of data through text generation, which ultimately leads to various communication methods changing, and in the end, there is a more ultimate direction. However, the regrettable thing is that there are still many people who only look at this. They say that what is good in terms of performance is Cloud 3 or Entropic Ramatu.

Today, it was in the news again. It's meaningless to continue testing things like the performance evaluation of L&M. Why did developers put the answer in the answer sheet for problems and keep testing based on that? Instead of testing with common knowledge, if we teach the answer by learning the answer, how will the performance indicators of L&M be? It will inevitably be much better than what came before, right? Because they already know the answer and just teach the answer, we no longer need to trust those parts of performance benchmarking tests. It's important. Some people need to continue with the LM part, that's correct. It's not saying it's unnecessary; it needs to be done. However, it's not necessary for everyone to only look at the basic model. Some people need to look at models that are tailored to specific domains.

Some people need to think about the parts that are related to what kind of sentences we will create, what kind of data we will use, and how we will execute this ultimately, and how our business activities are connected. However, there are people who need to think about these parts separately. But when someone with a very loud voice keeps talking, you might think, Is this relevant to me? I don't know what this is about. It looked pretty good, but I don't really understand. So, in fact, if you think that the LM or Chaeji PT generative AI you are currently watching is the same as the old phone shape we saw before, you should see that. Imagine how many people envisioned today's smartphones by looking at the brick phone and how it would develop in the future. It was surprising at that time, right? The fact that you could make calls outside without a cord. Do you remember that ad, Headquarters, come out, headquarters? Wow, it was innovative, but...

Is it just a fascinating world where mobility is just about moving without boundaries? No, what kind of companies make money from that? Services that create something to boast about when taking a picture, those are the ones making money. That's right. I think we are at the LM stage right now. No one knows how it will develop in the future or how it will be created. However, no one knows what direction the market is going. The debate over whether to use AI is over. It should be used. If wireless calls are introduced, there are many advantages over wired calls, so we have no choice but to go in that direction. There are of course people who say they will only use wired calls, but the trend is going that way. Nowadays, there are not many people with landlines at home. You probably know what that means. If it's confusing, what I mean is...

There are so many companies talking about generative AI-based systems. Hey, why don't you also do generative AI over there? Well, we also work with AI here. Look here. Of course, there are many duplicate companies. For example, they are divided into different areas, so it looks like this. Here, if you look, applications, services, models, infrastructure are broadly divided into these five categories. Even within these five categories, there are too many players. So, in this direction, it looks quite complex, but oh, those are companies that provide services. Here, companies develop applications, here is the platform model, and then the infrastructure. If you think about it simply like this, you can roughly grasp generative AI. So, in terms of models, Microsoft and Mistral, about a month or two ago, formed a partnership. In my opinion, with these L&M companies,

Platform operators and infrastructure operators holding hands, parting ways, and fighting each other will increase significantly in 2024. Of course, there will be more cooperative aspects rather than fighting now. Actually, these aspects will benefit consumers, but the crucial point is not about the performance of L&M, but about having partnerships that can encompass all of these things and integrate well with the company mentioned in the slide I just saw earlier. It is essential to have parts that can be well integrated with such partnerships. If you only focus on one part too much, even if the foundation model is too strong, there is no top or bottom. What should we do then? It's painful. That's why overall, such partnerships, tremendous partnerships, competition, and various mergers and acquisitions among them will be the key trend. And notably, what.

Microsoft said that's the mystery. So let's go back to the important part. Where are we going to use AI fundamentally? Please write down your thoughts. How has your life changed? There must be some changes. However, the crucial point is that AI understands to some extent what you say. You all know that by now. But are the apps and systems you use every day getting smarter accordingly? No. We are still pressing the order button the same way, searching for categories, and making choices in the endless UI, using apps that force us to make choices. Despite the era of AI being able to understand speech, we are still asking ourselves, What do I want to eat today? I had jjajangmyeon yesterday, what should I eat today?

I wish the calories were a bit lower and it would be nice if they could understand and respond when I talk about it, right? Do you think it's impossible? It's possible. However, those who are creating such services may not know because there is no relevant data yet. Due to the lack of sufficient data to create a service like that AI, there are not enough suitable data yet. The first person to do that could be Kakao or Baedal Minjok, or various other services. The opportunity to transition to such services has already begun. As I have been saying since last year, someone will run wildly, someone else will run wildly, and soon, within the next 2-3 years, or even sooner, there will be benefits through AI. This will happen where companies are concerned as well. We are not such a startup, and in fact, we have enormous data.

And already, where should AI be integrated with existing systems? What is the most important thing to consider with all these questions here? Where to focus on all the numerous questions, who will start writing first, how to safely use them, and what will be the ultimate penetrating content of countless questions with this? How can we secure more competitiveness, not with AI, but with my current and future competitiveness? The deep contents in the minds of decision-makers are not about the importance of AI. Whether it's AI or anything else, it's about how my business can continue to dominate in the present and future. But when it comes to that, if you keep going into that L LM LM, it's difficult to find the answer. So, in fact, AI is already being applied in all industries - legal, retail, medical, manufacturing - and results are coming out, and it is being applied in all sectors.

People can actually apply it, and that's why they look at it in six major ways. Well, what does it mean to enhance employees' experiences? It means changing all the experiences of employees, from basic system data to various customer education aspects. When we think about the biggest problems in manufacturing, if an accident occurs, it's often due to insufficient employee training. Yes, that's right. What's wrong with people? But they don't think that the problem could also lie in the training costs and methods. It means that employees can use new educational methods for accessing data or equipment. One of the biggest advantages of generative AI that many people talk about is hyper-personalization.

When creating a gamer travel program or providing travel services, it's not about group tours or programs for 100 people. Even with just 100 people, you can tailor the travel route to their preferences. It's possible to change parts related to customer engagement and business processes that were previously automated. You can use LLM to exclude certain parts, what was the bottleneck, whether you can use it or not. Therefore, in the most innovative parts of the RPA process automation process, what are the most innovative parts? It's the automation process itself. So, you have no choice but to go towards parts that accelerate innovation. And to accelerate innovation, you need to run fast. You need to run fast, but your shoes are heavy, your backpack is heavy, and there are too many things to carry. Can you run fast? No, you can't run fast. Then, what should you do?

You should start by implementing AI in the business that has been lightweighted to be able to run faster. Then, based on that, you need to reconsider the data a little more and proceed with the work from there. Otherwise, many companies actually end up being delayed in terms of innovation despite shouting for innovation. As for some representative cases, it's not so much about the names of the companies mentioned here, but more about how they have implemented changes. For example, KPMG has created internal AI agents to improve specific knowledge or areas within the company. On the other hand, for NHS, they have increased customer satisfaction by providing new services and continuous 24-hour service based on AI agents.

In the case of Symphony A, there are aspects related to financial crimes. Then, in the case of GM, they used Follow to expedite some business processes. Although not mentioned here, last year, they announced using Mercedes-Benz and GM simultaneously with Microsoft and GPT to introduce virtual agents into cars. There were many movements even in Korea early last year. However, instead of integrating it into the car first, they automated the processes and accessed the data within the processes based on LLM. They found that it was more effective to use it first. Of course, they will likely use virtual agents for consumers, but as they internally use these agents, it has led to further improvements, which is also a trend. This is from an IDC report, where it states that when companies invest $1 in AI, the return on investment is as much as 3.5 times.

There is a very interesting result. Now, to do AI, you need to pay for the API, invest in GPUs, and use the cloud. It was said that it costs money, but as it turns out, the return is as much as 3.5 times the investment. In Korea, there is a typical case. I'll tell you about it. Even though it's not in this data, a domestic mobile carrier provided a service to summarize and save call contents. They did it for phones that originally couldn't do that. Due to this and other services, some customers switched carriers. Looking at last year's accounting data, among the three carriers, the only one that saw an increase in revenue was that carrier. Isn't it amazing? Are you using a lot of AI? By using AI, companies can innovate services and show profits. I see it as a specific example. Of course, it takes time. It takes about 14 months for return on investment, of course.

This is because it's an average part, so it could be shorter or longer, but it took about 14 months. Another interesting trend from last year is that Microsoft's Copilot already surveyed users in North America and Europe. Surprisingly, 77% of users said they would not eat the free lunch given by the company every month, so they asked to use Copilot instead. Instead, because this also costs about $19 a month, $19 or maybe around $30, wouldn't the company spend 50,000 won a month on company dinners? I won't eat, I won't go to company dinners, so please use that money to get a Copilot Office account. It's a new welfare for employees. Yes, you can use it in the form of new welfare benefits. So, in a way, it means they don't want to do simple tasks, and in the future, this generative AA.

There are so many companies creating solutions related to this, which means that the market is opening up. Compared to last year, IT spending has doubled in 2024, and the investment in AI has increased. Looking ahead to 2025, 2026, and 2027, it is expected to continue to rise steadily. Predictions are being made that the parts related to AI within IT spending will continue to increase. Next, it is quite interesting to see the changes in demand for each industry. There is an interesting point that initially, there was tremendous interest in the legal industry. People in the legal field were amazed and shocked, saying, This is a must-have, but then realized the essence and said, Oh, this must be used, and then realized that it cannot be used legally. They said, Oh, this requires a person.

Suddenly, there is a drop in the industry. Of course, that would be a 'cassette' phenomenon. Due to the initial rise, the early adopters use it first, and many others watch and wait to see the results, leading to a period of observation. Therefore, there may be parts where many do not use it but watch. There are also continuous areas of interest, and suddenly, there is an increase in the parts related to Travel AI that was not considered very important. This is because there are continuous areas where it can be applied immediately. Because the tourism industry has been sluggish during the Corona period, the demand for tourism and related parts is increasing. In Europe, for example, as you may know if you have been there, group tour buses are blocking city entrances, preventing vehicle access altogether.

There is contamination by the sense of the group, various noises, and discomfort of local residents in the city. So, please make a quiet city. Cities like Vienna, Austria, are emerging, saying, Don't come in groups, don't bring group buses. How should we proceed? Who will create personalized routes that can create individual tourism programs? If 100 people are tourists, should each individual be taken care of? It's impossible. Who will do that? It will be done as I explained earlier, content curation. So, as I mentioned before, it's about creativity, but ultimately, it's a conversation. We know that AI can now have conversations. What's behind AI? Data. In other words, humans will finally talk directly with data.

Please let me know the sales data of the Sales Team 1 from last month compared to last year. It's such a great quality, right? But what should the person developing this do? Last year, they had to go into the SAP system, bring out the ERP, pull the data, and then analyze it. It will take a month or two. This is what they say when it comes to development. Why do you do it like that? You have to create a complex UI, with search, execute, select, and so on. Instead of that, you can just create a button, a dialog box, and do it as we did earlier, right? Please find relevant precedents in this document and let me know. It's so nice, isn't it? Is that impossible? Technically, it's not impossible at all. What's important is that all interfaces will change from now on. In the past, we lived in a world with only text, but now we live in a world with graphics, a graphical user interface.

In the era of mobile devices, where we carry graphics in our hands, it is now transitioning to the era of conversation. When we move into the era of conversation from mobile, many aspects can improve and change significantly. Earlier, I talked about Mercedes-Benz. If I'm in a car now, instead of saying things like lower the window or raise the window, it's not about that. It's more like, I feel stuffy and want to breathe some fresh air, please lower the window a bit. You need to speak like this. Just pressing a button and increasing the temperature by one degree won't do. So, in order to have such conversations, we need various sensor data, map data, weather data. For example, if it's too hot outside and you open the window, it won't work, right? We need to have conversations like this and move in that direction where AI forms that can understand such context will increase. It should go in that direction, not just about L&M.

There are too many parts that can be seen as conversational elements rather than functional elements, but in fact, there are many parts that we are concerned with. So, in the case of Mercedes-Benz, for example, there are enhanced voice commands, not just voice commands. And what's written in front of it, right? So, if you lower the window, how about turning off the air conditioner this time? Since I mentioned earlier to lower it, then it seems like the air conditioner should be turned off, right? We are having a conversation, not just one-way interaction, but interactive interactions can take place. And then, enhanced operations, followed by situational follow-up questions. As I mentioned earlier, depending on the action taken, there can be follow-up questions. In fact, because GPT, for example, or L, in the case of such chains.

You can link the contents of a conversation like a chain, which is an important function. Although you may not know what was said before, generally, in cases like this, you remember the previous conversation, so it is possible to have limitations on follow-up questions. Therefore, it is more important to move towards AI-based conversations rather than simply understanding what is being said. That's why I believe there will be an increasing demand for conversational AI. So, from the perspective of developers working on this app, what I always mention is that the apps we currently have are quite challenging in terms of interactivity. If you look at the mobile apps you are probably using or the apps from larger companies, you will understand this. However, as a result, the data is fixed, and every time there is an update or change, the cost and complexity increase. I was talking about a famous ERP.

I went to an ERP company and looked at the menu. When I saw this menu, it was so complicated. So, when you actually choose an option from the menu, you can select 20 different things as soon as you press the selection button. Customers keep demanding things all the time. However, since we have to make only one product, in order to respond to each customer, we created 20 selection criteria. Who is going to find those 20 criteria? So, you have to receive training. If you want to use the accounting program, you have to receive training to be able to use it. And if you receive the training incorrectly, you won't be able to use it. So, only those who have received the training know what to do. There is also a part about being stuck in a vicious cycle. Existing programs couldn't help it. Why can't people understand what others are saying? But now, the era has come where people understand each other. It operates based on natural language and is personalized.

I can provide services, and then because new features are not tied to the app, it's possible to naturally provide new features in the background. In the future, I think even shopping websites will just come out with a chat window. Bank apps, are they super apps? No, just having a chat window. Once you verify yourself, if you want to send 300,000 won to someone, you just say so, and it's done. It won't be an era of pressing complicated buttons. If you understand that, it's enough. It's not that you can't understand it. Also, many people have been worried about copyright issues and various aspects such as IP. Simplifying the issues of copyright infringement not only from the user's perspective but also from the perspective of I used this, but what if there's a copyright holder? These are some of the concerns, and in reality, the copyright...

It is not permissible to use the content of the person you have without permission, so it is essential for both creators and users to protect rights. Currently, many companies are doing so, but companies like Microsoft provide solutions for copyright issues that may arise from using their content. Responsible aspects of AI, like the six principles mentioned here, will continue to be emphasized. If these issues are not addressed, AI may become unusable. No matter how well it speaks, if it includes insults, vulgar language, or excessive use of profanity, it cannot be used for business purposes. Therefore, these aspects become crucial. In addition to open source application sources, slm.

So, it is expected that the importance of small language models will increase significantly. Therefore, I believe that they will be used a lot in environments without networks or for specific purposes. When you look here, instead of continuously pressing buttons on a kiosk to order a hamburger, going back and forth, you will simply say, Change one of the two burger sets to orange juice from cola. This era of ordering with simple commands like this will come. However, for that to happen, technologies such as reducing ambient noise of ordered data will be more revealed. Recently, related to these chipsets, you may have heard of NVIDIA's GTC event, where there was a long line. When Jensen Huang mentioned the new chipset called Blackwall, companies like Microsoft, Meta, Google, and AWS stood out.

All companies were busy competing to show off their partnership with NVIDIA by writing articles. Inevitably, for the time being, it is expected that NVIDIA will provide many advantages and features in that area. This is likely a clear fact. However, companies that create such solutions and infrastructures each develop and apply silicon optimized for their own environments. It is expected that such movements will continue to occur in 2024. Therefore, it is anticipated that the demand in the semiconductor market for such areas could potentially have a positive impact on the Korean economy. Not only in the semiconductor market but also in other industries, it is considered important overall to utilize generative AI in various aspects. Although I summarized it here, due to time constraints.

I will conclude my presentation with this without rereading the content. Thank you.

[D트렌드세미나] EP.1 | 글로벌 AI 트렌드와 전망 (이건복 상무)

Hanwha DREAMPLUS Official YouTube Channel