Hello, I'm going to give a little presentation now. My PowerPoint is working well. The technical foundation behind that, well, it's not really a technical discussion, but understanding that, in my opinion, AI is now being applied and utilized in various places. Many people are confused about what AI really is, or even artificial intelligence researchers themselves don't know or don't even think about it. So, if you think about what AI is, it's not easy to define clearly. At least since the 1950s, when AI started, the goal was to create a machine that thinks like a human. A machine that thinks and behaves like a human, typically a computer. What does it mean to think? In fact, we express thoughts in language. So, the most excellent thing that AI does now is to speak languages, language skills.

The most intellectual activity that humans engage in is using knowledge. Therefore, language ability is at the core of thinking. However, to think using language, one must understand the problem. In fact, perception is required. One must perceive the world, understand the environment, grasp the context, and perceive. Typically, human perception relies heavily on visual abilities, with about 80% of information obtained through vision. So, visual abilities are currently emerging, perhaps following language. Therefore, language abilities are currently emerging, and what AI is still not good at, and what needs to be done in the future, is action. Ultimately, creating an agent like a human being involves perceiving the world or environment, thinking, making judgments, making decisions, and taking action. That's what the original dream of artificial intelligence was, to create agents like humans or even robots. So far, this has been happening bit by bit, in fragments. So, let's go back to the beginning.

Because it's what we do best, how do we understand the surrounding environment? We understand it through words. Input is in words, thoughts are in words, and actions are in words. Therefore, we receive words, transform them into words, and generate words. So, what has changed now is that until recently, before generative AI, it only made judgments. It distinguished between this and that, analyzed things, but now, through deep learning, AI not only distinguishes but also creates. It accomplishes creation, writing text. In the past, when given a sentence, the question was whether it was grammatically correct or not, only testing and evaluating were possible. Now, it writes. Anyway, if we look at what AI is, it is about making machines do intellectual tasks like humans, and to do these intellectual tasks properly, understanding the problem is crucial.

You need to have the ability, then think about the results, and finally express them in action. Ultimately, by transferring it into action, one can become a complete machine-like human. However, even though AI has been studied for 70 years, in fact, it struggled a lot for the first 40 years. There was a well-known AI winter. The reason for this was that it was known from the beginning that in order to make machines intelligent, they need to have a lot of knowledge. Therefore, for decades, research was conducted to input knowledge into machines. The method of giving machines the knowledge that humans know is by programming computers. It was about coding the algorithms I thought of to the machines, then inputting the known knowledge through rules. Despite years of research on this, in fact, even with things like AlphaGo, they couldn't achieve that, nor with chatbots.

How to do it, people do not code knowledge to machines, machines learn on their own. So what artificial intelligence researchers have done is to study how to make machines learn on their own, which is actually machine learning. The field of machine learning itself has been around for a long time, but in fact, the foundation of this revolution ultimately lies in machines becoming smart on their own, not humans making machines smart. So, learning is important. If you look back to around the year 2000, you can see that artificial intelligence has transitioned from classical AI to modern AI. Just as physics transitioned from classical physics to modern physics with the emergence of quantum mechanics, AI is now transitioning from classical AI to modern AI. Therefore, it can be seen as before machine learning and after machine learning. Before machine learning, people programmed...

I initially tried to make the machine smart, and after machine learning, the technology now allows machines to learn by themselves, which is what machine learning is. Now, it has evolved into deep learning, and next is the current Generative AI. Through this, we have Watson, Google, autonomous vehicles. In fact, it started around 2011. This was the predecessor. About five years later, AlphaGo appeared, and then about seven years later, now we have GPT. It is still in progress. So, ultimately, the fact that machines learn is because they learn with data. So, the amount of data greatly affects the performance of artificial intelligence nowadays. The reason why GPT, language, came out first is that text data is the most abundant data. Since the advent of personal computers, all data has been digitized, and with the rise of the internet, all documents are digitized. Therefore, text data is the clearest and most abundant data highway. In a little while, I will explain it again, but anyway, that's why in programming.

The method of making machines smarter by rules has not made much progress, but recently, if you look, by collecting a lot of data and using computing power to teach, you can see that artificial intelligence problems are being solved faster than expected. So, when you look at the subject, what is a learning system? Right now, artificial intelligence is smart, but what's important is the learning system. At this moment, the intelligence is increasing. That's the most different thing. In the past, computer programs were developed by developers, and the intelligence was at the moment of creating the program. But now, it's not like that. The program goes out into the world, and humanity is actually still teaching the ultimate intelligence of all humanity. Still, why? Because OpenAI people are still tuning the ultimate intelligence in some way.

It seems to me that it's not such a simple issue in terms of language problems. So, in generative AI, what is it learning? Just like I mentioned earlier, until recently, it was good at distinguishing between A and B, giving pictures and guessing if it's a dog or a cat, distinguishing between the two. Generative AI goes beyond that. If you train it with dog pictures, it will generate new dog pictures. If you feed it text and let it study Shakespeare's works, it will write a diary that Shakespeare might have written, using Shakespeare's style of sentences or words. In that sense, generative AI is doing an amazing thing, actually creating things. Especially with language, since humans do most intellectual tasks like work based on language, by reading and writing, so we presented on all business aspects just a little while ago.

Artificial intelligence can now be utilized in almost every field, especially in tasks like work. Even though it may take a short time, you really need to understand what learning is to truly understand artificial intelligence. Of course, you don't need to research everything, but it is really important to know what learning is in order to understand what AI is good at and what it is not. Learning is just like how we humans learn. It's the process of trying something, gaining experience from past experiences, and trying to do better in the future. So, basically, it's about trying, making mistakes, and learning. For example, the two-legged doll here wants to learn how to walk far. How can it do that? It tries something, falls, learns to avoid it next time, tries something else, and eventually learns to walk further. Humans learn to walk in about a year.

It takes a few minutes to learn this through simulation on a computer. So, after repeating it 40 times, it immediately falls over. If you walk a little further and accidentally repeat walking further, you walk a little more. By repeating this process, you eventually learn to walk without falling over. Look at the collected data and write a speech for me. At first, you write a nonsensical speech. If, for example, you write three PowerPoint presentations, among them, which one do you like more, if someone evaluates it once, imitate the better one. Then write three more speeches. Among the three, if you let me know which one you like more, I will imitate the good one and write a new speech.

Writing like this, in a way where humans and machines collaborate, is important. There are still tasks that humans do, so it is essential. Therefore, what is better, what is more important, what has value, these are still human tasks. Evaluating these things is something machines still cannot do well. In fact, there is a method called early-stage feedback, which evaluates what is better for machines. For example, teaching dishes what is good or not good. If you teach just a few, the rest will be done tremendously by the machine. Tasks like composing new speeches can be done more diversely than we think. That's why, based on the same principle, whether it becomes a form like a person with two legs or a form like an animal with four legs, learning can be done in a similar way. Even getting tripped and getting up again can be learned with the same principle. Therefore, learning is a process of trial and error.

It's a mistake, then correcting an error, and then it's a system that continues to improve itself. The most important thing is a system that continues to improve itself. So, if we talk about learning to distinguish between dogs and cats specifically now, this is supervised learning. Usually, we teach the correct answers. We provide pictures, teach that a dog is 1, and a cat is 0. Humans only do the work of providing, for example, 2,000 dog pictures and 2,000 cat pictures, teach for a year, and then leave the rest to the machine. So, the machine now learns. For example, in the algorithm of deep learning, which is a deep neural network structure, we stack layers of neurons like the nerve cells in our brains. Therefore, deep learning gives input, calculates output, but initially, it will make mistakes. Even if you give a dog picture, it may not answer 1 but 0.6, for example.

So, a 5th-order was created within 1 - 0.6 0.3, and we correct the 5th order. The one written in blue as yn 1, FX n 0.6, and the correction method is to provide input, calculate the output, calculate the error, correct it, and this is now the backpropagation algorithm, which was developed back in 1986, over 50 years ago. The algorithm underlies everything we are doing now. Of course, it has become more complex than this, and has been engineered and scaled up, but the basic principle is all here. So, input, calculate the output, correct any errors, input again, calculate the output, correct it, and if you write it out in a formula, we can move on. So, input, calculate the output, correct the error, that's why what people do is define errors, then continually modify the structure of the neural network in the direction of reducing errors. We keep changing the structure of the neural network to define errors in the direction of reducing errors, based on what we see and learn in our minds and brains.

It can be seen as an artificial intelligence algorithm that mimics mathematically. The advantage of this is, as I mentioned before, if I have 20,000 pictures of dogs and cats and can't do well with 20,000 pictures, if I gather 200,000, it will do better. Even if it performs well, it will do even better with 2 million pictures, even better with 20 million pictures, 200 million pictures, 2 billion pictures, and 20 billion pictures. If you continue to increase the number of dog pictures, performance will continue to improve. Now, the performance is absolute, of course, depending on how difficult the problem is, but in any case, the important thing is that as the machine gains more experience about the world by continuously collecting data, the performance improves. This was something that was impossible with the old programming methods. So, by gathering a lot of data and continuously learning with computing power, the performance keeps improving. That's why deep learning is fundamentally based on the principle of learning.

So, in the past, when neural networks had many layers, it was too complex at this stage, so we couldn't train them by directly inputting dog and cat photos. In the old days, computing power wasn't good, so we had to develop image processing programs using dog photos, for example, and extract only features like eye shape, ear shape, and mouth shape. Then we would gather those features and train them with simple machine learning algorithms. But now, with the improvement in computing power, in the era of deep learning, we can just give a whole cat photo like the one you saw earlier, and even if we say to do something with the whole thing, we can do it. For example, with GPT, even if you give tens of pages or a whole book, it reads them all and learns something. So now, the complexity is tremendously high.

It has increased, but the principle is similar. So, if you look, in 2014, Facebook started recognizing whether a person in a photo is male or female. It's not easy to determine if it's a man or a woman, even people can't do it 100%. If they wear makeup or something, it's hard to tell if it's a man or a woman. Therefore, by giving input photos and processing them through a complex neural network structure using a deep learning approach, the machine produces output. When you see it, in the front, it seems like the machine is recognizing human faces, but as it progresses, strange patterns that are completely unrecognizable to human eyes start to appear. However, the machine combines all those facial features and creates its own language, eventually distinguishing whether it's a man or a woman. AlphaGo utilized that. What you are seeing now is a video processing deep learning structure called 'Convolutional Neural Network (CNN),' and AlphaGo used it as it is.

Instead of a human face, a board is given in place of a human face, and the output is not 0 or 1 but binary coordinates of 19 by 19. It is a bit complex to decide which move to make at the current position. AlphaGo used reinforcement learning overall, and within reinforcement learning, it had to predict how good a certain action would be. It used deep learning in a supervised learning manner to predict how likely it is to win in the future depending on the move made. It's a bit complex. It's different from just learning English. When you make a move, it involves predicting how good it will be in the future. Therefore, it is challenging. Anyway, in the end, it underwent training, and then, now, as I mentioned earlier, even AlphaGo, like what you saw just now, is not generative AI. Because it distinguishes whether it is better or worse to make a move in Go.

However, what does generative really generate? The principle of a sample is all here. Instead of distinguishing between dogs and cats, if you input a dog picture, it will also output a dog picture the same. So, if you give a dog picture, it generates a dog. But here, as you can see, the number of neurons is gradually decreasing and it was slow, so we compress it. In fact, this is what humans do. We look at all kinds of information all day long, but we mostly discard it, compress it, keep only the important things, and store it. So, based on the concept of compression, if there is complex information, we encode the input, compress it, and then create an image again based on those features. If you provide a video as input, it creates a video. Similarly, with GPT, instead of dog pictures, it was given text. If you give a speech, it will generate the same speech. While memorizing, it also learns things like grammar in the language.

If you describe it in Korean, it means learning the grammar and meaning of Korean in PowerPoint. If you describe it in English, it means learning English. In fact, PowerPoint doesn't care what language is, it just creates some meaning from various combinations of words. So, this can be seen as generative. The most amazing thing about GPT, which is currently available, is that everyone in the world knows difficult words. GPT is a very technical term, pre-trained transformer, a very difficult word. Pre-training is preparing in advance, so it has studied all the documents in the world. And transform is a mapping function that converts into a new structure. It is a deep learning structure, which means it is just a more complex version of this. So, when something comes in, it transforms and continues the transformation.

When it comes to generating output, it is considered as creating, not just distinguishing. Machines can clearly understand this fact through machine translation. Google already had a generative AI that significantly improved the performance of machine translation in 2017. So, if you provide the English sentence I love you as x and the Korean sentence 나는 당신을 사랑해 as y, and input many corresponding corpora for training, you will see that it involves a deep neural network structure with many layers. Through encoding and processing information via deep learning, it learns the characteristics of English sentences, encodes the information, and then converts it to generate Korean sentences. By decoding, when you provide x as English and y as Korean, it becomes a machine translator. If you provide x as Korean and y as German, it becomes a Korean-German machine translator. In fact, depending on the data provided, various translations are possible. NVIDIA has developed a new MA.

Learning method generating UN and realistic-looking faces using GAN. Now, the key is knowing and discerning new 사람 얼굴 (human faces). By training with many human face photos, this is a learning model called GAN. Even if you encode it like this and see it as similar, it's all fake photos, whether it's a man, a woman, or a different race, creating similar photos by wandering around almost the same photos. The fact is, instead of pixels, GPT generates characters. If you think about it, it's the same as generating words. Of course, now that language has an order, it looks at the left and generates to the right. Of course, with sentences, well, for example, but in the case of pictures, it looks around and generates the current pixel. It generates that point according to the context around it, and that's the only difference. Basically, that's why recently, they call these things foundations, foundation models, and that's why by putting a lot of text and making a large language model, like GPT, that's what it is.

After creating a foundation model, you can later apply it to various fields such as finance or law. Then, you refine or transfer the learning before applying it. In this process, artificial intelligence mimics the way humans learn. Just like when we accumulate general knowledge in elementary, middle, and high school without deciding on a major, learning languages, mathematics, and history, it is like building a foundation model. It's just about studying a lot. Then, when you go to college and decide on a major like law, you fine-tune by learning legal terms, adding another round of training. After graduating and entering the workforce with a law major.

I went into the finance field, then fine-tuned it back into the finance field for applications. So now, there are super large AI models, like the large language model. But now, not only with text but also combining text with images and various things, now multi-modal models are being developed. So, artificial intelligence is just beginning in my opinion. With a lot of data, you can do something, which has given a lot of confidence. However, data is still lacking in most fields. In fact, data is still lacking outside of text. So, when that accumulates, it continues to progress towards something that can truly be done like a human. Now, more and more computing power is needed. As you can see, large models and super AI are getting bigger. It seems to be increasing linearly here, but due to the logarithmic scale, in fact, it is exponential. It's tremendously huge.

Now, the trend is that computing power issues are becoming more and more of a capital fight. So, recently, if you look, not only can machines generate text when given text or generate videos when given videos, but if you look at machine translation, it can handle various languages like English, Korean, Japanese, and more. By learning text and images together, in the future, when you provide text as input, the machine can generate images, and conversely, if you input an image, the machine can generate a description in text. So, you can see that this is a machine-generated image. Next, as you saw in the video earlier, recently, AI has been able to generate even long videos. In fact, research has been ongoing for short videos for the past 2 to 3 years. So, recently, these things are happening, such as Cheetah and what Google is doing. One important aspect is that machines can generate code. Machines can generate code, which is programs, computer programs.

It may seem difficult to people, but in fact, it is easier for machines than Korean or English. That's why the words have very clear meanings. Our natural language changes depending on the context, but ultimately, there are languages that are easier for machines. This is why there is a lot of interest in using simpler languages for machines. In fact, coding education should now move in a quite different direction. Recently, there are companies that only provide codes, and even with the emergence of machine learning, a significant part of coding automation is being achieved. Especially now, as programs are generating code themselves, just like generating text, it is possible to collaborate with machines to write programs much more easily. Therefore, in the past, in the era of AI, in computers, there was machine language, like assembly language.

I talked about advanced languages and created another language on top of that. People no longer need to learn the language of machines, it's almost the same. Just give a PowerPoint presentation in English. You don't need to learn programming languages like natural languages. When we speak to machines in a language we know, they understand, write code, write text, create reports, draw pictures, design, and so on. Of course, programming languages will not disappear now. Still, to handle very detailed matters, there need to be people who can program computers and use very complex languages. However, in many cases, when things are digitized, machines now do a considerable part instead of ordinary people. Therefore, especially the fact that it becomes generative has a tremendous impact. The reason is that it really writes text, draws pictures, and so on.

I'm designing and doing this kind of thing, so the future of work is really changing now. Artificial intelligence that uses human knowledge is trying to automate certain tasks. A lot is happening through language right now, especially in handling language. Therefore, just because that part has come, it doesn't mean there's nothing for humans to do. I don't think so. Humans still have the ability to judge, which machines don't have yet. There are no criteria for values. What is more important and what is more interesting, these things are not yet possible for machines to determine, so the intention and such aspects still need human guidance. If you try to do everything up to that point, you will lose in competition. Leave that to the machines. So, having a good assistant or an intern beside you, thinking of it as teaching, will allow you to do more work.

It seems right to try to do more important and valuable work effectively. As you know, there are drawbacks. The current drawback of the technology is that optical character recognition is excellent. It's amazing. What's even more amazing is that it learns only from text as if it were human. Have you ever thought about it? Optical character recognition has never seen a mouse here. In fact, it has never touched a mouse or tried to use a computer with a mouse like this. Instead, it manipulates the computer with a mouse. I've read all kinds of articles about the mouse. It's a bit ironic. So, it's natural not to understand 100% and not to understand. Therefore, hallucinations and illusions are inevitable, and in fact, hallucinations, in my view, can happen even to humans. Maybe.

I don't know the essence of intelligence because it's somewhat about imagination, so that's why hallucination, depending on the terminology, may sound negative, but it's somewhat about imagination. That's why it's necessary evil, and that's why talking about things that are not true is a bit of a difficult issue now. Technically, trying to do things that are not possible requires an enormous amount of data and energy, and still, humans need to provide some feedback to some extent. Without us knowing, in fact, AI is teaching everything. Actually, it's not explicitly stated, that's why philosophers realized this early on. Pragmatic philosophy, for example, already exists in the world, and we need to figure out what to do, but now, with the AI world, it is separated, and only things that are digitized are used for learning. For example, only using text for learning, this is what we don't understand, so in reality,

So, the phenomenon of hallucination, if you look at this picture, do you know why hallucinations occur? I trained a model to generate that picture only with animal photos. But later, when I input human photos and ask it to recreate based on that prompt, since it only learned from animal photos, it interprets everything as animals. Therefore, hallucinations can occur because my experience of the world is limited, so with that limited experience, I can only perceive the world as that. For example, if someone doesn't have the common knowledge that Einstein is not Black, they could generate such a picture. The way to prevent this is to interact with the world, not just through text or photos, but by truly interacting with the world, experiencing it, and turning that experience of the world into data for the machine to learn from, so it can truly understand.

What we are doing is using our bodies, having physical senses, collecting data from them, and learning from it. This may be a way to overcome it. This is now the future. Up to now, there are still a tremendous amount of things that generative AI can do. To address such shortcomings, research is being conducted, as you can see here. In the future, when robots are actually asked to go somewhere and fetch water, if it's about water, robots cannot help it. But as the robot arm goes, recognizes the water, actually picks it up, and runs errands, through this process, as robots gain experience by doing errands, phenomena like hallucinations that we saw earlier will decrease. Next, this is the Robot Cup, an annual competition held by the AI Society. It took place in Bulgaria last year, and although it was a scene where we won, if you look at it, it depicts home.

Robots can be helpful in the environment, understanding facts, engaging in conversations to a certain extent, receiving tasks in language, executing them, and providing useful services. For example, in scenarios like serving people at events, robots need to grasp and serve. Due to time constraints, I will stop here. Something more interesting you might have seen is using a humanoid robot with a body to experience the surrounding environment through the five senses. As AI becomes smarter, Tesla has declared that they will deploy humanoid robots similar to humans in Tesla's car factories in about 3 to 5 years. These robots will walk like humans, carry things like cars, and even assemble them by bending down. The way of teaching them is different; when humans perform actions, they are digitized, and robots mimic and learn from them.

That's what we call imitation learning, where people teach machines by imitating human actions, so it can mimic what humans do. The advantage of a humanoid robot is that it can imitate what humans do very well because it has a human-like finger shape. Lastly, this is a video that came out just a week ago. It's from a humanoid robot company called Ure AI, which collaborated with OpenAI. The OpenAI GPT robot understands this and explains it as if it understands the world like humans do. It describes what it sees at the moment, like giving an apple, understanding intentions quite well, asking questions about thoughts, perceiving the environment, thinking, translating into actions, and expressing. Of course, this is still a demo video. These things are starting to happen now. So, to summarize, artificial intelligence actually has about three main aspects.

I tried to directly input the past data, information, and knowledge, and now the next step is for machines to have wisdom. I tried to input knowledge directly, but actually failed. What I am doing now is providing data and enabling machines to learn knowledge by themselves. It's not about forcing knowledge into children but rather allowing them to accumulate experiences by taking them to museums and playing with them. This is the current successful method, but the drawback is that there is still a gap with the world, so in the future, it seems that one must go out into the world, physically gather data, and learn.

[D트렌드세미나] EP.2 | Next-Gen AI 트렌드와 전망 (장병탁 AI연구원장)

Hanwha DREAMPLUS Official YouTube Channel