[Lee Myung-ho's Future Preview] AI Language Ability and Human Writing Skills

AI Natural Language Processing Capabilities Show Rapid Improvement Last Year
Writing Program 'GPT-3' Enables High-Level Topic Conversations

Since AlphaGo's victory in the intellectual game of Go, considered the pinnacle of human intellect, in March 2016, what groundbreaking artificial intelligence technology can be identified? AI technology is already utilized in many products and services we use daily. Video services operate AI recommendation algorithms, and when typing on smartphones, automatic suggestions for words or sentences are based on AI technology. The technology that once distinguished between dog and cat photos now finds and classifies faces of oneself and friends in pictures. Although autonomous vehicles are not yet commercialized, they are rapidly advancing. AI technology is used in 80% of advanced products and services. Among the AI technologies developing across various fields, some are emerging as potential game changers.

Among the remarkable achievements last year, I focus on the rapid improvement in natural language processing capabilities. Particularly groundbreaking is the AI writing program GPT-3 developed by OpenAI. GPT-3 attracted attention by demonstrating human-like writing skills. On a chat site used by 30 million people, the GPT-3 bot posted hundreds of writings over a week without being suspected of being AI-generated. Most comments written by the bot were neither biased nor harmful. In one post on the topic of suicide, it even wrote the following, stating that thinking of supportive parents made it give up on suicide: “What helped me the most was probably my parents. I have a very good relationship with my parents, and no matter what happens, they have always supported me. I had many thoughts of suicide in my life, but because of such parents, I never attempted it.” The conversations extended beyond everyday topics to high-level subjects such as the existence of God and the fate of humanity.

AI Language Programs Used in Business and Writing Fields, Expanding Their Scope Steadily

GPT-3 can write long texts when given a starting sentence and topic. The Guardian even published a long article written by GPT-3. Although GPT-3 demonstrates human-like writing skills, including ethical issues, it has not yet officially launched as a service and is still under development. Currently, GPT-3 can process a staggering 175 billion parameters. Its memory alone reaches 350GB. In 2019, the GPT-2 version processed 1.5 billion parameters, and GPT-4 is expected to handle trillions of parameters this year. GPT-3 requires enormous hardware performance and is expected to start service with limited licenses due to concerns about misuse.

GPT-3 is reportedly also being developed in a Korean version. Korea has long developed natural language processing programs and has skills comparable to winning scholarship quizzes. However, it is still considered not on par with GPT-3. If Korean language processing technology falls behind, foreign entities could lead Korean writing programs. Foreign programs are also considered ahead in Korean translation. Society is based on language. More cooperation and investment from the government and companies are required for Korean language processing technology.

AI language programs are already used in business areas. Although many nuances of human language pose obstacles, AI chatbots have already shown satisfactory results in customer service. Recently, their use has expanded to summarizing complex research reports, analyzing financial documents, telephone marketing, and call center records.

However, AI writing programs still face many challenges. As the AI chatbot “Iruda” was embroiled in controversies over sexual harassment, minority discrimination, and bias, leading to service suspension after just three weeks, it is not easy for AI to possess sound common sense and ethical awareness and use language like humans. Current AI natural language processing technology is based on deep learning that learns from vast amounts of human-written documents. It arranges words and sentences by following the most frequently used patterns in human-written texts. Such learned algorithms are difficult to apply to other domains, and their reasoning and knowledge are opaque, making it hard to explain why certain sentences were written. They imitate human sentences but do not possess human-like knowledge, common sense, or reasoning abilities. How to equip AI with these abilities remains a challenge in natural language processing.

AI Entering the Stage of Conscious Thought, Preparing for Human-AI Collaboration

Generally, human thinking is divided into ‘System 1’ and ‘System 2’ thinking. Daniel Kahneman, a Nobel laureate in economics, described System 1 thinking as intuitive, fast, and automatic. Distinguishing cats from dogs, recognizing words on passing billboards, or completing the phrase “Romeo and 00” are examples of System 1 thinking that require little conscious processing. AI programs have shown excellent performance in many tasks that humans can process in less than a second. System 2 thinking is slower, more analytical, and involves deep conscious thought. It is used when reasoning is needed to solve abstract problems or handle new situations. This includes solving complex puzzles, proposing new concepts, planning, understanding causal relationships, and determining the appropriateness of specific behaviors in social contexts. Our efforts to understand the world, create a better world, think, and acquire new knowledge belong to the System 2 thinking domain.

Currently, AI language processing is at a stage attempting to transition from System 1 to System 2 thinking. To enable System 1 thinking, AI has been trained using supervised learning methods with well-structured datasets. Recently, unsupervised learning models that mimic how humans learn about the world through open exploration and reasoning have been developed. It is uncertain whether AI will achieve System 2 thinking in a few years or this year. Just as AlphaGo suddenly shocked us, when AI begins to think like humans, we need to prepare for how to accept and use it. Even with System 1 AI handling simple repetitive tasks, we feared the disappearance of human jobs. Before a new shock arrives, we need to prepare by understanding what true human abilities are and how humans and AI will collaborate.

Lee Myung-ho, Planning Committee Member, Yeo Si Jae Foundation