Kwon Dae-seok, chief executive officer (CEO) of Hyntel, is a computer scientist and entrepreneur who has researched "strong AI" that surpasses human intelligence.
He founded the supercomputing solution company Clunix in 2000 and wrote the "Manifesto for Strong AI" in 2016. In 2022, he established the company Hyntel, which aims to develop a "master algorithm" capable of self-modification and self-evolution. Hyntel is a shortened term for "hyper-intelligence."
The reporter thought of him because forecasts and warnings that AI could fundamentally change modern society have been coming in succession. Futurist Ray Kurzweil has advanced the anticipated arrival of the singularity from 2045 to 2029. Geoffrey Hinton, the father of deep learning, urged political intervention in preparation for AI potentially initiating wars.
Jensen Huang, CEO of Nvidia, which recently topped the market capitalization rankings, noted, 'In less than five years, AI will likely outperform most people by about 8% on math and reading comprehension tests, as well as exams for doctors and lawyers.'
CEO Kwon is more radical. He believes that strong AI reaching the singularity will emerge within two years. His 'Manifesto for Strong AI,' created in 2016, begins with 'Happiness is a specific internal state of our mind and feelings' and concludes with 'AI should not threaten humanity's survival and happiness.'
What exactly are strong AI and the singularity, and what impact will they have on humanity? I met with CEO Kwon of Hyntel and CTO Peng Won-ki, who are researching AI algorithms at a shared office in Seongsu-dong, Seoul. Recently, the two unveiled a paper decoding AI with a transformer architecture using quantum mechanics. The following is a question-and-answer session.
─ The reason for establishing Hyntel separately.
Kwon Dae-seok = “In 2014, I was lying down due to an ankle injury, and suddenly thought that the AI algorithm I had envisioned 24 years ago could now be implemented due to sufficient computer memory. That’s when I started my research. Two years ago, after receiving private equity investment, I was able to step back from managing Clunix and founded a research-focused corporation.”
─ The direction of research and development at Hyntel.
Kwon Dae-seok = “There are two aspects. First, to significantly improve the currently widely used GPT model and create a 'post-GPT.' Second, to develop the AI model 'SANC (Self-organizing Active Network of Concepts)' envisioned during my undergraduate studies. Since SANC is not based on deep learning, it does not require large-scale matrix operations. Instead, it possesses capabilities such as additional learning, partitioning learning, and real-time feedback integration. Its lower computation requirements result in reduced learning time and operational expenses. I believe combining post-GPT and SANC can create 'strong AI.'”
─ What exactly are strong AI, AGI, and the singularity.
Kwon Dae-seok = “Strong AI is an AI that excels in solving all problems a person can tackle. It is commonly referred to as artificial general intelligence (AGI). An AI that only solves problems in specific areas, such as Go like AlphaGo, is classified as "weak AI." The singularity refers to the point where no human can understand the advancements in technology. As strong AI continually undergoes self-modification and self-evolution, humanity will reach a level of singularity intelligence that is difficult to comprehend.”
Peng Won-ki = “CEO Kwon believes that AI should have self-awareness, but I'm not so sure about that. However, I believe AI with self-evolution capabilities can be implemented.”
─ Why are strong AI and singularity intelligence important.
Kwon Dae-seok = “Singularity AI will solve problems that require intelligence beyond that of humanity (nuclear fusion, cancer treatment, quantum computing, room temperature superconductors) more quickly. Each solution could give rise to corporations valued higher than Samsung.”
─ With the continued dominance of OpenAI, do you have a chance of succeeding?
Kwon Dae-seok = “Seeing the capabilities of 'ChatGPT' made me feel frustrated to the point of depression. However, since it is based on deep learning, ChatGPT has many issues. ChatGPT was created by training a supercomputer worth hundreds of billions of won for several months. The learning time is lengthy, and it requires a lot of data for training. It also incurs massive expenses for operation. This is because it consumes substantial computing resources each time it provides an answer. After AI learning is completed, adding specific knowledge or correcting errors is also difficult. Most importantly, it is impossible to discern why it produces certain answers.”
─ Hyntel claims to have partially understood the operating principles of AI, which used to be like a black box.
Peng Won-ki = “OpenAI's ChatGPT and Naver's HyperCLOVA X operate on the 'attention-based transformer' technology developed by Google in 2017. Transformers are neural network technologies that track the relationships between data such as sentences and words to learn context and meaning. The operation of transformers was similar to the minute steps of 'Feynman path integral' used in quantum mechanics to predict when a subatomic particle transitions from one state to another at a specific time. This suggests we have found a clue to solve the black box problem of deep learning. The research content was made public on the paper-sharing site arXiv (https://arxiv.org/abs/2405.04620).”
Kwon Dae-seok = “Feynman path integrals are used to calculate the probability of a very small particle's state changing from one point in time to another. By applying the Feynman path integral to transformers, we could compress long information into a few state vectors. This means minimal loss of contextual information without using a large amount of memory. I believe this provides a way for corporations without vast capital to develop and operate AI.”
─ AI models such as ChatGPT, Gemini, Lama, HyperCLOVA X, and Mistral are flooding the market.
Kwon Dae-seok = “Ultimately, it’s a contest of 'who reaches the singularity first.' In this game, finishing second doesn’t mean much. In fact, the method of AI implementation isn’t that important. The first AI to reach the singularity will accumulate self-modifications, widening the gap with the second-place competitor. Google, which lags behind OpenAI with AI performance at the second level, is likely to be anxious. At this pace, I believe singularity AI will emerge within two years. The rate of technological advancements is very steep, and researchers are quickly finding the mathematical and engineering definitions of intelligence.”
─ The aftermath of reaching the singularity.
Kwon Dae-seok = “In the past, capital and labor determined competitiveness and value, but in the future, it will be who has the better AI. The hegemonic advantage of nations possessing strong AI will be clear. All of humanity could fall into a state of unemployment. The foundation of capitalism, which relies on value creation through labor and distribution through wages, will be shaken. Therefore, a new distribution system that does not depend on wage labor will be necessary.”