Future of AI: Trends, Machine Learning, and Language Models

AI is transforming industries, making understanding its trends essential for future business planning.

### The Evolution of AI: Understanding Machine Learning and Language Models
* 🤖 In the next 30 years, significant industry changes driven by AI are expected to **mirror the mobile technology revolution** that occurred 15 years ago, emphasizing the **importance of understanding AI trends** for future business planning.
* 📈 The 2025 AI Trends Course has been developed to educate corporate members on the technological evolution of AI, highlighting the **foundational role of machine learning** in these advancements.
* 💡 Despite being considered an "old technology," machine learning remains **crucial and widely used**, although its complexity often leads to a **lack of awareness about its applications**.
* 🔍 The discussion will focus on how AI, particularly **machine learning and L&M (Language Models)**, operates, its **limitations**, and its **current and future utilization** in various industries.
* 🏫 The speaker emphasizes the need to **revisit the basics of machine learning** to better analyze and predict future developments, particularly as L&M continues to **evolve and impact various sectors**.
* 💻 Companies are already utilizing AI in practical applications, with examples including a university hackathon where **16 out of 20 teams incorporated AI**, showcasing its **current relevance and utility**.
* 🌍 The narrative draws parallels between the transformative impacts of the internet, mobile technology, and AI, suggesting that while AI is currently surrounded by **hype**, it is poised to bring about **real and substantial changes** in how we interact with technology.
* 🔄 The conversation concludes by stressing that **effective data management is essential** for successful machine learning and L&M applications, as **poor data organization can lead to significant operational challenges**.

In 30 years, just like mobile technology 15 years ago, we are facing significant changes in the industry designated by the emergence of AI in 2025, occurring in a 15-year cycle. Without understanding the transformation that AI products and services will bring to customer relationships, we cannot anticipate what next year will hold. Planning next year's business without knowing AI trends is meaningless. Therefore, we have prepared the 2025 AI Trends Course, which clearly outlines the technological evolution of AI and the changes it will bring to the industrial landscape, along with the points that corporate members need to consider.

Ultimately, it all originated from machine learning, right? Isn't it an old technology? No, it's not an old technology. It has done significant work and has served as a foundation for further advancements. However, is it truly useful? Yes, it is useful and is still being used extensively. But why are we not aware of it? Because using it is complex. The data input must be handled very carefully, and overall, it is a sophisticated system, like a car.

When we talk about it, it’s almost passive. Still, L&M is all about it; they excel in certain areas. Yes, we are continuing the analysis of companies by a developer with 30 years of experience in Season 3 of AI. So far, we have mostly covered the latest topics. However, originally, when we planned this, we intended to start from the basics of machine learning and AI, as we need to have a clear understanding to analyze and predict the future. We had intended to focus on the fundamentals, but there have been so many interesting stories that we kept postponing it. This time, we really need to solidify the basics to discover new stories, don’t you think? Please give us a lesson today. Alright, so today we will discuss how AI, specifically machine learning and L&M, operates, its current limitations, how it is actually being utilized, and what the future holds.

It would be nice to take some time to thoroughly examine whether this works. When we look into these matters, we can interpret the trendy discussions we've had before. We can see what is happening, but I think it would also be good to cover some basic topics. One thing we need to clarify is that you are someone who already has a solid foundation, and you are currently working on a project using L&M. While working on this project, you are also conducting a lot of research for the Park Jong-cheon 30-Year Developer Series. You are analyzing companies and directly engaging with those who develop and use L&N, so you have access to real-time information that is both the latest and foundational. I hope you can provide a class that builds a solid foundation in this regard. Yes, from that perspective, I have been directly involved in L&M development and have been engaged in various activities, including a recent one.

Two companies held an AI hackathon, and the results were quite good. I saw how a typical software company can transform with AI. There was also a university hackathon where 20 universities gathered to actually develop software. Out of those 20, almost ten teams utilized AI. We both went there as judges. In this hackathon, the students were expected to have a basic understanding of front-end development, and back-end development was also essential for creating a service. However, it turned out that leveraging AI was already a basic requirement. It wasn't specifically an AI-focused hackathon, but among the 20 teams, 16 of them had already incorporated AI. So, when I see what I have created and how companies are actually trying to use it, it is very exciting.

There are many examples of AI being not just a future concept but something that is already here beside us. I was reminded of this once again. Among the comments we received, some people said, Isn't AI just an outdated technology? It's a technology that will disappear like blockchain. It's only being used for trivial things. However, as you just mentioned, there are already many applications in the way we develop and work. 

To talk about AI and L&M, I always prepare a line: the internet really changed the world. We can't go back to how things were before. Mobile technology changed things too. Just like that, AI is finally going to change the world. When the internet first emerged, I don't think anyone could have imagined what we would do with it. Initially, it was the same with mobile technology; now, when we take the subway, we use it in ways we never thought possible.

When I was using the phone, I couldn't imagine traveling around the world, pulling up something like Google Maps. We may not fully grasp it yet, but based on my experience and the results from many companies utilizing this technology, it's clear that it's already happening. You can be confident about it. Speaking of the internet, let me share a brief story from my youth. I first accessed Yahoo when my cousin showed it to me. Before that, we were using PC communication services like Naoli and Hi-Tech. When I entered Yahoo, I asked my cousin, What is this? His only explanation was that PC communication was something local, while the internet connected you to the whole world and was in English. In fact, the functionalities available were not much different from what we could do with PC communication back then. But now, the internet has evolved.

PC communication is completely different, right? But when we think about L&M, we are trying to do only what we can within the current standards. So when we ask, What is this? we say, It's something that computers do. However, once L&M matures to a certain extent, it will be doing things that we can't even imagine. I think it would be good to reflect on that future. Just now, we talked about the internet, mobile, and AI, and what's interesting is that there are a lot of AI bubbles. If we think about why there are bubbles, AI has been around for quite some time, but now it's all the noise around LLMs and ChatGPT. If we think about it, while the internet made a huge impact, it was actually the World Wide Web that made the breakthrough. And although mobile made a huge impact, it was the iPhone that really succeeded. In the end, the leaders in technology and those who commercialize the technology are different.

그래서 AI 체스는 세계 챔피언을 이겼지만, 실제로 이를 상용화하는 것은 아직 이루어지지 않았습니다. 그래서 지금까지 안 된 이유는 그 기술이 없었던 것이죠. 결국, 웹과 아이폰이 세상을 바꿨다면, 이제는 LM이 세상을 바꾸고 있습니다. 정확히 말하자면, 인터넷과 모바일, AI가 아니라 웹과 아이폰, 그리고 LM입니다. 어떻게 보면 제품이 세상을 바꾸는 느낌이네요. 그리고 웹 이전에도 FTP나 유즈넷 같은 많은 것들이 있었지만, 우리가 사용하기에는 너무 어렵고 불편했습니다. AI도 마찬가지로 머신 러닝이 굉장히 좋지만, 생각보다 사용하기 어려웠습니다. 하지만 LM, 즉 챗GPT는 자연어로 작동하니까 정말 쉽습니다. 그래서 이렇게 폭발적으로 성장할 수 있게 된 것입니다.

I think everyone needs to go all-in on timing. We shouldn't worry too much about AI laws; we need to prepare for the future. Let's learn exactly today. You mentioned LM earlier, right? You've probably heard a lot about what LM is, but could you start by explaining the basics? 

The interesting thing is that we've been talking a lot about LM, but there was also something called ML, which stands for machine learning. However, people tend to overlook machine learning, so it would be good to understand it. Isn't it an old technology? No, it's not an old technology. Ultimately, we are doing our work using computers. If we're not using computers, then LM or ML are actually useless. In the context of doing work with computers, one part is done directly by a person. For example, a carrot...

Items are all listed in the market. You can look at them and say, This is a counterfeit product. By observing, you can also identify where it is sold in Gangseo District. It's something that a person can do, or it can be done through algorithms. For example, if you search and see the word Gangseo, you know it's in Gangseo District, and if you see Nike, you know it's a shoe, and so on. Machine learning can also be utilized. Machine learning is trained to make judgments like a human. It excels at classification. So, machine learning can be used, and L&M is a more advanced version of machine learning. For instance, if someone misspells Nike as Nai Ti, it can still understand contextually that this refers to a purchase service for Nike. Therefore, it can work with that information, of course.

When it comes to algorithms, they are the best. People accurately write down their ages, and if everyone does this in Korean, then the algorithm can work 100%. However, if that’s not the case, we inevitably have to use artificial intelligence, and in that situation, we need to rely on machine learning or LLMs. Unstructured data is much more challenging to handle. This is why machine learning and LLMs are necessary. Most of the data that exists in the world is unstructured; structured data is better suited for LLMs. That’s right, machine learning can work to some extent, but LLMs perform much better. It’s quite difficult with just algorithms. But then, isn’t it enough to just learn LLMs? Is it really necessary to learn machine learning as well? However, there are situations where machine learning can work, and it’s cheaper and more efficient in those cases. So, LLMs ultimately emerged from machine learning. Let’s briefly touch on what machine learning is.

It would be good to build, but the first point is that traditional programming involves inputting data and rules, which then produce an output. For example, if you input 1 and 1 with the rule of addition, the output is 2. Then, if you input 2, the output is 3, and if you input 2 again, the output is 4. This was very simple. However, since we had to calculate every possible case, there were limitations, and exceptions to the rules kept arising. So, despite continuously updating the rules, we couldn't find the correct ones. This led us to question why we were trying to find rules. Instead, we could reverse the process: by inputting data and output, we could derive the rules.

Can you create it? So if we go back to the numbers we just did, 1 plus 1 is 2, and 2 plus 1 is 3. What do you think the rule is here? Ah, I can figure out that this is addition by myself. Just like a human, a machine can do the same. It rolls the dice a million times to find out that this is addition. This is a coincidence. So, it finds a rule and tests it out: does it work or not? And while doing this alone, it discovers addition. Now that it has found the rule, if we give it 2 plus 2, it can determine that the result is 4. Humans do this, algorithms do this, and machine learning does this. When we talk about algorithms, we set the rules. It can be very complex; for example, if male is checked, it means male, and if female is checked, it means female. If it says not specified, then it could be male.

Even if you set rules like this for men, if exceptions occur, it doesn't work out. So, conversely, people respond like this: Oh, this was a man, or This was a woman. If you try to find the rules, then they find the rules themselves. So, if it says man, it's a man; if it says male, it's also a man; if you write it in Chinese characters, it's a man; if you write just ni, it's still a man. It's about finding the rules on their own. Then, if there is a lot of data, better rules can be found, right? For example, if you give one picture of a cat and one picture of a dog to train, it will match those two pictures. But if you give a different picture, it might not be able to match it. If, by chance, the second picture is very similar to the first one, then it can be quite ambiguous, and it might not match.

When it first came out, Google said it was fascinating to upload photos of chickens for free, and many people did so. After I uploaded the photos, it suddenly started categorizing them. For example, this is a soccer scene, and this is a wedding. However, there was an error at that time that caused quite a stir globally. There was a Black couple, and it was categorized as a gorilla, which caused a lot of uproar. So, the service was quite problematic. Yes, that's right. It's exactly like that—classification. When you provide a photo and give the correct answer, the system learns from that. For instance, you give a lot of gorilla photos and some photos of people, as well as soccer photos. But there will be a limited amount of data given. As a result, it may not have been sufficiently trained with the data, which can lead to incorrect classifications in such situations.

So here, for example, if we take 1 and 1, it equals 2; 2 and 1 equals 3; and if we ask what 2 and 2 is, our goal is to find the rule. The reason we find the rule is to solve the next problem. We give the cat calculations and train it a lot, and what we want is to present a new picture and see if it can guess it. However, it might be able to guess the picture or it might not. The interesting part here is that if it can't guess, we can provide the correct answer and retrain it. So if we come back to the question of what 1, 1, 2, 2, 1, 3, and 2, 2 is, it might figure out by itself that this is addition, and it could say that this is the answer. However, we also introduce the concept of parameters here, which is an important concept in machine learning. Parameters can be various combinations of rules within this framework.

In mathematics, there can be one, two, or even three symbols involved. So, if there is only one mathematical symbol, we add. But let's assume that there are two mathematical symbols. In that case, the first symbol could be multiplication, and the second symbol could be addition. For example, if we start with 1, we multiply it by 1 and then add 1, which gives us 2. Then, if we take 2, multiply it by 2, and add 1, we get 3. So, the initial 1 plus 1 gives us 2, and 2 plus 1 results in 3, satisfying the condition with completely different parameters. Then, 2 multiplied by 2 equals 4. However, it turns out that was incorrect; it should have been 2 equals 5, indicating that the rule was wrong. When I tried again, I discovered that this involves multiplication and addition. Therefore, when there are two parameters, one can be addition while the other might not be present at all. This means that from the beginning, we are dealing with two parameters.

There were two correct rules, and if we consider it as one parameter, there was one correct rule. However, when we introduced two parameters, we were able to interpret cases where it became five. This means that when we were given pictures of cats and dogs, if we set the parameters to, say, ten, we were able to classify them correctly or incorrectly. But when we were given new pictures of cats and dogs and it was wrong or right, we could say good job and so on. In this ongoing process, the existing parameters might have been insufficient. The more parameters we have, the better it is. So, we can't strictly say that parameters are the rules, but the methods are different. However, in general, one might think that the more parameters we have, the more rules we can consider. Would that be a reasonable way to think? Yes, yes. So, generally speaking, when we talk about machine learning,

If we talk about parameters, let's say there are ten thousand or so, when it comes to LM Pro, it transitions to a different level. This level is fundamentally different. The more parameters there are, the more diverse issues can be addressed, which is beneficial. We tried to find a rule like 1 plus 1 equals 2 and 2 plus 1 equals 3, focusing on parameters and determining whether we are adding or multiplying and then adding one, and so on. In this way, we fed all the information available on the internet into the LM for it to learn and be able to respond. Conceptually, machine learning and LM are the same. There were parameters in machine learning, right? Yes, indeed. Because we input a much larger amount of data, had significantly more parameters, and trained it for a much longer time, it has become LM, rather than just appearing out of nowhere.

It's not about technology. That's why we need to understand machine learning, right? So, it's good to have a little knowledge about it. In fact, people don't have much understanding of it anyway. Anyway, it finds rules and has parameters. Just knowing that is enough. When we drive a car, we know there’s an engine and that there are four wheels. It's similar to that. 

Now, machine learning is based on large models, right? I've heard the term machine learning before, but it never seemed to have this much impact. I think I’ve seen articles that mentioned distinguishing well, but I wonder why it has become so important now. We need to look at the stages of artificial intelligence. In the past, algorithms were the way to go. Even in chess, where the world champion was defeated, all the rules were created and incorporated. It wasn't machine learning back then.

Then, the person who created that must be really good at chess, right? They made all those countless chess rules. But if you add exceptions to it, it becomes foolish. So, chess has its limitations. However, Go has so many combinations that it failed to establish rules. Before AlphaGo came out, even I couldn't play Go. Right? So, AlphaGo succeeded by learning on its own. The algorithm based on those rules had its limitations, but when it learned by itself instead of having humans create the rules, a completely new world opened up. In the case of cats and such, even if we try hard to distinguish them by examining their ears and noses, there are still strange variations in Africa.

If you bring in a cat, it will be incorrect. Then, if you code that, it will also be wrong, leading to an infinite loop. Ultimately, you have to provide an enormous amount of cat photos and an enormous amount of dog photos for the machine to learn. So, it has now learned. It's the opposite concept of what humans used to do, and we made machines do it. That's why it's called machine learning; literally, it's the learning of machines. In the end, there needs to be a lot of data. In the past, it was impossible to find thousands of cat photos or thousands of dog photos, right? That's true; back in the days when there was no internet, it was hard to gather that data. That was the problem. Secondly, to train this, the computing power needs to be very good. You need to use a lot of NVIDIA GPUs. The learning algorithm that splits these cat and dog photos for training has only been developed relatively recently; it wasn't something that existed a long time ago.

So, as these learning algorithms emerged and a lot of data became available, and since there are many GPUs in the cloud, machine learning, which had been just an idea, finally started to take off. This laid the groundwork for the emergence of systems like AlphaGo, and now we even have models like ChatGPT. Google offered to open their services for free, encouraging people to upload photos to train their models. Ultimately, this required a vast amount of data, and they began to collect all of that. The reason language models (LM) came about is that there was enough data accumulated on the internet. If there had only been a little data on the internet, LMs wouldn't have been possible. The trends of the outputs reflect the collective knowledge that humans have accumulated so far. Similarly, current LMs are based on even more internet data and utilize powerful GPUs for highly effective learning algorithms.

It has come together to create L&M. I noticed that this topic is also mentioned in the book you wrote. It talks about big data, and it's not just about having data; we also need to label it for machine learning. We need to determine whether it is a dog or a cat, and then train the model accordingly. This kind of information has accumulated even more, right? So, one of the points being made is whether this will continue to develop in the future. NVIDIA GPUs are already quite good, but will there be even better ones? There will be, although there may be some limitations. Still, I think there is room for further advancement. Although we have used a lot of data on the internet, more data can still emerge. Data from around the world will continue to come out, improving the quality of data. Naturally, learning algorithms will also get better, so L&M will be even more advanced than it is now.

Will we develop? Yes, this is a very important point. For example, the current LM as a finished product has its limitations compared to what we want to achieve. However, if the current L&M is not a finished product and can improve, we should create work while thinking about a better future. Even when looking at the three components, the algorithms are continuously being improved. As you mentioned, there is a vast amount of data that has not yet been digitized. Since the emergence of LMs, AI technology has improved, which helps a lot in digitizing analog data. For instance, OCR technology has become so advanced that you can just take a picture and it works. In the past, we had to scan each item individually, but now, as analog data is increasingly being converted into digital data, the amount of good data will continue to grow, just as you mentioned.

NVIDIA's GPUs are continuously improving, and it seems they can only get better. Moreover, other companies are also developing competitive products, which is expanding the market even further. Additionally, social data is not being utilized to its full potential right now, so the more we use language models (LMs), the more data we will incorporate. This creates a positive feedback loop where people will generate higher quality data. While the pace of development has been rapid so far, I anticipate that the speed of progress will accelerate even more in the future. If we saw advancements from GPT-3 to GPT-4, I expect even more astonishing results as we move from GPT-4 to GPT-5. Now, I understand that LMs are continually improving. As for machine learning, you mentioned it earlier, but I wonder if it is used consistently beyond just Google Photos. Is machine learning being completely replaced by LMs?

In reality, this is very good, but ultimately, ML has played a significant role ahead of LM, and it has laid the technological foundation. However, we need to examine whether this has been truly useful. It has been useful, and it is still being used a lot. But why don't we know much about it? This is much more complex. It is complicated to use. Earlier, when we talked about the internet, we mentioned web and FTP. Many people may not know what FTP is, but it was a way to download files. Nowadays, we simply download from the web, so it's not necessary anymore. When we received the address, it seems that FTP was attached when we downloaded, right? It was quite complicated. So, the internet has both FTP and the web, but the web is easier to use, which is why it has succeeded. Similarly, ML is the same.

Writing was very difficult. As I mentioned earlier about the cat and dog, I had to organize cat photos and dog photos, and I needed to label them as cat or dog. It required thorough preparation. Now, if I had just thrown it out there and asked, Is this a cat or a dog? that would be one thing. However, with machine learning, it's not like that at all. For example, you might just be given two documents and asked which one is more polite, and the model would determine that. We need to be very careful about the data we input for machine learning, and we also have to be very selective about the data we use for writing. Overall, it’s a very sophisticated system; if we compare it to a car, it’s almost manual. Ordinary people probably won't use it because it’s difficult to write with. So, even if they are within that system, for example...

If it's an online shopping mall, then there's a recommendation engine within it that suggests new products based on people's purchase history. This purchase history needs to be well organized, and based on that, machine learning is trained so that when you ask, the results come back. But this process itself is very complex. You don't really need to do any of that; it's all part of the engine that was developed during the learning phase. For example, if I have been to restaurants for a month and then I say, Plan my restaurant visits for this week, it will organize it well for me. 

When comparing usability between ML and L&M, it feels very inconvenient. It's like receiving car parts, assembling them, and then having to drive them. On the other hand, with L&M, it's like getting into an Uber that just comes to pick you up. It’s much more convenient since I don’t even have to drive. I think it’s that easy. So, are people using it less now? Well, still, yes.

Earlier, I mentioned that LM is very cost-effective and performs well in specific fields. Therefore, there are still areas where it is used, and I believe those areas will continue to thrive. Before the advent of LM, I identified about four fields. The first is recommendation systems, like those used by Netflix. The second is time series forecasting, which includes weather prediction and stock price forecasting. The third is image recognition, and the fourth is natural language processing. 

While I was working on natural language processing using ML, LM was invented. Initially, I was focused on tasks like translation and summarization with ML, but then LM came along, and everything started to change. Now, LM has essentially dominated this field. Whether it's recognition or prediction, LM can handle it all. However, in practice, natural language processing is where LM truly excels and demonstrates superior performance. In contrast, the other fields are still well served by ML, which is also cost-effective, so we still use it.

It doesn't. Can LM do that? Well, it's expensive, and it might not be able to do it. In the case of ML, for example, with image recognition, I need to provide it with pictures of dogs and cats for training. Then, it will generate a model that fits my purpose perfectly. After that, if it encounters something it can't recognize when I give it a photo, I can provide the correct answer and retrain it. However, with LM, we are using something that has already been created. So, for example, if I give it a picture of a cat and ask whether it's a cat or a dog, it should be able to guess. But there are cases where it can't get it right. In that case, there is no way to correct it. That's why when LM first came out, if it made a mistake, there was no way to fix it. So, we need to explain it differently. We have to handle it separately, which is a hassle since we need to prepare everything for ML, but we can modify it as we like.

There are certain products that are somewhat expensive but convenient to use, and they cannot be changed. Therefore, from the company's perspective, when providing a service, it is essential to fix any bugs immediately if certain functions need to be implemented. In that case, machine learning might be a better option. It's good to look at the field in this way. 

Moreover, AI can perform a lot of tasks that humans do. However, it doesn't take on every single job. For example, while a car is a very useful tool, it ultimately serves the purpose of getting us from one place to another. Of course, you can watch movies, sleep, or go camping inside the car, but AI can also do a lot of things. It can handle various tasks effectively. 

What humans have discovered includes things like recommendations from platforms such as Netflix or Coupang, as well as stock price predictions and weather forecasts.

However, this is really difficult. It's still a field where we haven't had much success. I think it's a bit challenging in the area of language modeling. So, we are continuing our research in this field, but it hasn't yielded significant results. The next area is image recognition, and image recognition is doing really well. Companies like Google have already pre-trained a lot of models. Even without us inputting data, it performs well. In fact, we can take the existing models and add our additional data to further train them. So, personalized recommendations and image recognition are areas where machine learning is performing excellently and can be utilized. Many companies are still using machine learning for this, so there might not be a need to necessarily transition to language modeling. We need to start retraining language models from the kindergarten level, while machine learning has progressed quite a bit, and there are already students who have learned up to high school education. With just a little bit of additional training, they could use it even better.

It's easier to customize for the necessary purpose. Customizing LM is not easy. The four items you just mentioned were actually discussed at the beginning of our Season 3. Now that some time has passed and we’ve tried using LML, it makes more sense to us. Yes, yes, yes, this is very important. Ultimately, AI is about where humans can assist. There’s personalized recommendations, then a bit of time-series forecasting, then recognizing something through text, voice, or images, and finally, natural language processing. These six areas are continuously relevant and very important. Interestingly, the first three have a significant impact, but the last one, natural language processing, is much more disruptive. So, if this allows us to drive far in our cars, it means we can reach places we haven't been before.

In that regard, as we processed nature with ML, we created LML, and when we tried processing natural language with LM Pro, we found its utility to be extremely high. Now, we have surpassed the initial three applications and are achieving dozens of times more utility in the L&M field. Many companies are utilizing just this one aspect of natural language processing. Companies like Netflix, Coupang, and many others are creating and using these systems, including Carrot Market. Initially, everyone developed their own systems, but over time, companies like Amazon and Google have created systems that simply provide APIs as platforms. For example, in terms of recommendations, Amazon has a ready-made recommendation API. So, if I have an online store, I would have product information and customer information, and these customers have purchased these products.

This is about providing information on people, products, and interactions between them, which can then predict what someone might buy next. If you ask the person, they will also provide insights. That's why we have effective recommendation engines. Additionally, in areas like image recognition, Google offers Cloud Vision AI, which has been trained on a vast amount of data, making it easy to use. You can provide an image and it can identify whether it's a cat. If we get it wrong, we can provide a new picture to help it learn more. There are great platforms available, so we can simply use them. This process is known as fine-tuning, where we adjust a pre-trained model to better suit our needs. Fine-tuning makes it easier to modify the learning. For images, while we have trained well on text, we can provide additional data for further training. However, this is just one part of the process.

It's not about ending with a number; it's a system that continuously operates automatically. For example, in Coupang, when recommendations are made, this learning happens, let's say, once every hour. So, if I made a lot of recommendations for an hour but people didn't buy anything, then that needs to be communicated. This way, it learns that people didn't purchase. Conversely, if people do buy, it should learn from that as well. This process keeps changing. For instance, if today I make some recommendations and suddenly a drama airs featuring a popular necklace that everyone starts buying—something they previously weren't interested in—then the recommendation system quickly learns from that data and adjusts accordingly. It's a very real-time and fast response. However, this doesn't happen with language models. So, does machine learning become smarter as it accumulates more experience? Yes, it does become smarter.

There are various factors to consider, but you can follow that trend. For instance, if it's a shopping mall, without doing anything special, in the summer, you would recommend summer items, and in the winter, you would recommend winter products. In 2023, you would suggest items that were trending at that time, and in 2024, you would recommend based on what’s popular then. This naturally follows the trends. So, everything moves according to that data. As the data increases and accumulates, you can predict more and more, right? While the system improves, there are limitations to this as well. If you input too much data and train it excessively, it can become costly. For example, if there are ten million users and ten million products, would you train the system based on what those users purchased over the past week or over the past month?

Are you going to train the data from the purchases made over the last three years? If so, the amount of training and the cost will change. However, some people bought something three years ago, and I haven't bought anything in three years. Then, they come to shop today. If we only train with one month's worth of data in between and not the three years, this person wouldn't know anything. On the other hand, if we train with three years' worth of data, it would be too expensive. This makes things very complicated, so there are limitations. We can't train with too much data. Therefore, it would be good to quickly grasp and follow the current trends. I see. Here, we've emphasized data a lot. If it's an online shopping mall, then the data needs to be well collected and managed, and that needs to be inputted into the machine learning system. So, the ML platforms would be something like Amazon or...

Google provides great support, but this data is unique to my company. Therefore, managing that data well on a good data platform is very important. Yes, data can become a competitive advantage. Right? Most companies that attempt to use machine learning and fail do so because they haven't organized their data properly. For example, in an online store, there are products and customers, and these customers have purchase records. If these connections are not well established, it can lead to issues. For instance, if a person has bought a product three times, and when you check the product, it appears to be empty. It was originally supposed to be a t-shirt, but due to poor data organization, it was listed as a t-shirt three times, yet the entry is just blank. Even so, they managed to sell products somehow, despite not having organized the data correctly. However, when you try to use that data for machine learning, it becomes unusable.

In this case, if something like this happens, it will lead to incorrect results. It will provide misleading information, right? Yes. However, the same applies in the era of LM. If we want to work with LM, we need to have data, but if that data is not organized... For example, if we give some messy documents and ask to create a report, then it becomes impossible to organize them. So, whether it's in the era of ML or LM, having a good data engineering process and a solid data platform is very important. Developers often mention the saying, garbage in, garbage out. This is also true for AI. Recently, companies like DataDog and other data-related companies are thriving because they need to gather all that data.

"AI의 미래에 있어 LLM이 전부가 아니라는 것 아시죠?" (30년 개발자 박종천)

Technology(기술), Trend(트렌드), Transformation(혁신)에 관해 
꼭 알아야 할 지식.  

일하는 사람들의 지적 동반자 티타임즈TV. 

www.ttimes.co.kr

Contact
주소  서울시 종로구 청계천로 11(서린동,청계한국빌딩)
e-mail  ttimes6000@gmail.com
Tel  02-724-7718

티타임즈TV

The Impact of AI on Industry Transformation

AI is set to bring significant changes to customer relationships and the industrial landscape, with its emergence in 2025 following a 15-year cycle.

Machine learning, often seen as an old technology, is still highly useful and complex, serving as the foundation for AI advancements.

The integration of AI in various projects and companies is already happening, with its applications becoming a basic requirement for success in software development.

The Role of L&M and Machine Learning in Technology

L&M is currently operating within existing standards but is expected to evolve and do the unimaginable in the future.

The AI industry has experienced bubbles, with LLMs and ChatGPT causing recent noise, while the true breakthroughs in technology have come from the World Wide Web and the iPhone.

Machine learning, often overlooked, plays a crucial role in the work done with computers, and L&M is an advanced version of machine learning, excelling at understanding unstructured data and providing contextual insights.

The Process of Rule Discovery in Programming and Machine Learning

Traditional programming relies on inputting data and rules to produce an output, but this approach has limitations due to exceptions and the complexity of rule-setting.

Reversing the process by using data and output to derive rules is a more effective method, similar to how both humans and machines can discover rules through experimentation.

Machine learning further enhances this process by using large amounts of data to find better rules and allowing for retraining when incorrect classifications occur.

The Role of Parameters in Mathematics and Machine Learning

Mathematical operations can involve one or two symbols, with different combinations leading to varying outcomes.

The introduction of more parameters in a process can improve classification and rule determination.

Machine learning models with a larger number of parameters can address diverse issues, and the transition from traditional algorithms to self-learning models has significantly impacted artificial intelligence.

The Role of Data and Computing Power in Machine Learning

Machine learning relies on vast amounts of data, such as cat and dog photos, to train models effectively.

The recent availability of powerful GPUs and advanced learning algorithms has accelerated the growth of machine learning, leading to the emergence of systems like AlphaGo and ChatGPT.

The continuous improvement of data collection, labeling, and digitization, along with the development of competitive computing products, is expected to further advance machine learning and language models in the future.

The Role of Machine Learning and Language Models

Machine Learning (ML) has laid the technological foundation for progress, but its complex and manual nature makes it less user-friendly than Language Models (LM).

ML is still widely used in areas like recommendation systems, time series forecasting, image recognition, and natural language processing.

While LM has dominated natural language processing and offers convenience, it can be expensive and less adaptable for tasks that require retraining or correction.

The Role of AI and Machine Learning in Service Provision

AI and machine learning are essential for fixing bugs and implementing functions in expensive, convenient products.

While AI can handle various tasks, it doesn't take on every job, with machine learning excelling in personalized recommendations and image recognition.

Companies like Amazon and Google are providing ready-made APIs for these services, making it easier to fine-tune and modify pre-trained models for specific needs.

The Continuous Learning of Recommendation Systems

Recommendation systems, like Coupang's, operate in real-time, learning from user behavior and adjusting recommendations accordingly.

The system's effectiveness is influenced by data accumulation and trend recognition, but there are limitations in training with excessive data.

Proper data organization and management are crucial for successful machine learning, as messy data can lead to incorrect results and ineffective AI.