**The present and future of AI virtual humans seen by Kakao Brain**
*When will artificial intelligence (AI) virtual humans that think and speak for themselves be developed?*
On the night of the 27th, Kakao Brain held a real-time talk session under the theme of ‘Digital Humans and AI Virtual Characters’ in the Kakao Talk Open Chat Room Voice Room. Kakao Brain CEO Il-doo Kim, Vice President Park Byeong-eun, and Business Development Team Leader Kim Jae-in participated as presenters.
In this session, it was predicted that AI virtual humans could come out as early as three years.
CEO Kim said, “According to the conversation partner, an AI model that thinks and answers directly is expected to be created within 3 to 5 years at the earliest. AI that reads math problems and converts them into coding to solve problems is already available. Here If we develop a little more, we can reach it quickly.â?
In order to create an AI virtual human that can interact and talk with humans, a language model trained with large-scale data is essential. Therefore, research on advanced language models at home and abroad is in full swing, but the current level of technology is insufficient. AI that speaks according to a set scenario occupies the majority of the market. AI call center employees and bank tellers are representative examples. If you are in a situation that deviates even a little from the scenario or you have to think complex, the AI â??â??cannot properly respond. It means that it is easy to speak according to the script, but it is difficult to set the mood while having a conversation. Emotional exchange is also impossible. An advanced language model is essential to create interactive virtual humans. Large-scale data needs to be trained, and a large amount of capital and time is invested in this process. However, the opinion of Kakao Brain is that there is a shortcut. Kakao Brain focused on the process of children learning a language.
“Children learning to speak, similar to machine learning learning process… Expected to realize human-like AI models”
CEO Kim said, â??(I have a 24-month-old child) I didn’t know what I was saying, but I learned (language) by guessing the next word. I don’t know what it means, but it connects clothes with pink. Also, if you tell them ‘car is blue’, you don’t know what blue means, but when you say car, blue immediately pops out of your mouth,” he said. In other words, he said that it was just the stage of the ‘next talk generator’ that creates the next word well.
CEO Kim said, “At some point, a child started to know what ‘blue’ was. From then on, when he saw blue, he began to say the word blue. After experiencing these experiences, if you look for brain science or any intelligence growth well (AI language model) I thought that there might be a contact point with learning),â? he added.
Vice President Park, who is raising two children, also added an opinion. First, he said, “When the second follows the first, there is a very similar phenomenon.”
He continued, “I don’t know what I’m saying, but when my sister says something, I follow it. Even if I don’t understand what she is saying, I just imitate the sound. If I repeat the sound and see a similar word in a book, I attach it (or not) and notice it.
Team leader Kim also said, “Children don’t know, but don’t they talk first. Then, while talking, they think about it againâ?¦ Just as they learn the concept more and more after that, the same goes for computers. You have to understand that’, if you proceed in this way, I hope that eventually you will become a human-like computer or model.”
“Digital human communication method, it is not different from virtual influencers or real people in the online environment”
It is true that the market value of virtual humans itself is great even if it is difficult to communicate with humans.
CEO Kim said, “Virtual humans such as Sua are already active in the marketing field. This is a typical marketing case. The intelligent aspect is also important, but there are still many areas that can be done with only the appearance of a virtual human Even if it is not in the marketing field. “He said.
Vice President Park said, “Now, a technical environment that is visually closer to reality has been prepared, and various platforms and social media where digital humans can act as ‘content’ are actively being operated.” He continued, “The fact that a lot of services, including Instagram, can be used also has an impact. Apart from this, isn’t it already playing a role as a digital human, such as a banker?”
Team leader Kim also said, “When promoting a company, it seems that we communicate more actively through virtual influencers. Don’t we do a lot of communication with real people online, too. Being a digital human is not different in the way we communicate.” .
He added, “There is an expectation that virtual influencers such as ‘Rosie’, who have been working hard recently, will be able to communicate better with people if they expand their scope further.”
Kakao Brain challenging the ‘unthinking question’… Concentrate on research and development of super-giant AI model technology
Kakao Brain is an AI technology R&D company headquartered in Korea. The company’s mission is to challenge the ‘unthinkable question’.It came from the intention to inspire innovation so that everyone can live a better life. Kakao Brain has developed a number of AI services and models that focus on improving people’s quality of life, such as minDALL-E, KoGPT, and CLIP/ALIGN.
Recently, Kakao Brain also released an upgraded version of Mindali, the image generation model (RQ-Transformer), to the open source community GitHub. This model, consisting of 3.9 billion parameters, is a ‘text-to-image’ AI model that has learned 30 million pairs of text and images. It is characterized by significantly improving image quality while reducing computational cost and speeding up image creation.
Compared to Mindalli, the RQ-Transformer model has tripled the model size and doubled the image generation speed and training dataset size. In particular, in the case of Mindali, it was close to reproducing ‘DALL-E’ released by ‘Open AI’, an American AI development company, but RQ-Transformer is different in that it was developed with Kakao Brain’s own technology.
https://www.ajunews.com/view/20220429075611583