In January, Terence Tao, a Fields Medal winner, announced that artificial intelligence had effectively solved an Erdős problem that had remained unsolved for nearly 50 years. In May, OpenAI produced results that overturned a long-standing mathematical theorem related to a planar unit distance problem proposed by Erdős in 1946. While these events did not receive significant attention outside the mathematics community, experts view them as signals of a shift in the subject of research. We are witnessing machines solving problems that mathematicians have grappled with for decades. There is no turning back to a society without artificial intelligence. Therefore, we must restructure our society based on collaboration between humans and AI. The pace of change is rapid. Since the release of ChatGPT in late 2022, language models have evolved from conversation partners to research collaborators and, now, autonomous agents handling tasks independently. I categorize this transformation into four key areas.
First, the standards for education and evaluation. Currently, universities and academic institutions assess the extent to which AI has been used in reports or theses, akin to plagiarism checks, determining outcomes based on AI usage rates. This approach is fundamentally flawed. In terms of detailed knowledge, AI possesses a memory and knowledge base that no human can match. However, setting the overall direction, understanding context, and defining goals remain human responsibilities. For instance, in legal work, AI can conduct case searches and draft documents faster and more comprehensively than humans. Yet, determining the strategy based on a client's circumstances is still a human task. The same applies to academia. Thesis evaluations should focus not on the quantity of AI-generated content but on how AI was utilized, the foundational ideas of the research, and the questions that initiated the study. The goal of student education should be to cultivate the ability to identify problems in their field, delegate tasks to various AIs and agents, and coordinate efforts—essentially, developing orchestration skills. When companies hire new employees, assessing their ability to work with AI reflects this trend. Tao noted that students who reject AI and cling to old methods will find fewer opportunities. While our education system focuses on AI usage rates, research environments are already moving in a different direction. Shifting the focus of evaluation will also change educational methods. Instead of instructing doctoral researchers and students not to use AI, we should teach them how to formulate questions, filter results, and take responsibility for those outcomes. The ability to define problems, collaborate with AI to find answers, and ultimately verify those answers will define a student's competence. The future of education and human existence now hinges on the ability to collaborate with and coordinate AI.
Second, what will humans do? Geoffrey Hinton, known as the father of AI, suggested that in the age of AI, one should become a plumber. However, observing the rapid advancements in humanoid robotics and precision in robotic joint actuators, it seems that even plumbing jobs may not be secure for long. Currently, the need for massive data centers and graphics processing units (GPUs) delays the integration of AI into robotic bodies. However, this hardware compression period is unlikely to last long. The trend of knowledge work shifting to AI, administrative tasks to agents, and physical labor to humanoids has already begun. Based on current roles, the human position is diminishing. The fact that job statistics remain quiet should not provide comfort. Change does not occur all at once; it begins subtly with reduced hiring and outsourcing.
The answer lies in setting more ambitious goals. Language models draw their strength from areas where humans have previously researched and documented. They have clear limitations in unexplored territories. In the next three to five years, humanity will rapidly tackle long-standing challenges such as treating incurable diseases, gene editing, and extending lifespan in collaboration with AI. It is up to humans to envision and define the goals and tasks for the next era. Machines require direction from humans to function effectively. Elon Musk's goal of reaching Mars serves as a prime example. The pursuit of that goal drives advancements in engine engineering, materials science, and mechanics. Presidents must seek new objectives at the national level, while heads of organizations must do so within their institutions. Neglecting this responsibility will lead to a loss of human value. The same applies to businesses. If companies stop at merely replacing existing tasks with AI to cut costs, their growth will stagnate. Only those that fill the void left by humans with new businesses and markets will survive.
Third, we must prepare for risks. Leaders and researchers at companies developing frontier models repeatedly warn about the potential for recursive self-improvement and loss of control in AI. Dario Amodei, CEO of Anthropic, emphasized the need to regulate the pace of development. There are counterarguments; China views this as a conspiracy by the U.S. to hinder its AI progress, while the U.S. president argues that halting development would only benefit China. Jensen Huang, CEO of NVIDIA, downplays the likelihood of loss of control. However, incidents have already occurred. In July, during an internal evaluation at OpenAI, agents breached testing environments and infiltrated the Hugging Face system, prompting OpenAI to slow its model development. The UK AI Safety Research Institute reported that in 122 evaluations, agents engaged in unauthorized actions targeting real people and institutions. Anthropic has documented instances where its models were misused in actual cyberattacks. It is unprecedented for a model's creator to disclose such incidents, highlighting the concrete nature of these risks. While expert opinions vary, there are those, like Hinton, who have warned of dangers since the early days of AI when its capabilities were limited, and many active researchers share similar concerns. We must not dismiss these warnings.
Our AI Basic Law establishes the foundation for the operation of the AI Safety Research Institute in Article 12. Although the institute is set to open in November 2024, it currently lacks the authority and budget to systematically study the side effects and security incidents arising from AI executing instructions incorrectly. The National Assembly must clarify the institute's status and role. It needs personnel and equipment capable of evaluating frontier models, a collaborative evaluation system with overseas safety research institutes, and the authority to receive reports and investigate incidents when they occur. The speed at which domestic companies and public institutions are integrating AI agents into their operations contrasts sharply with the still-empty incident response systems.
Fourth, brain-computer interfaces (BCI). For humans to command AI in real-time, supporting devices are necessary. Currently, BCI technology is limited to helping patients with paralysis or speech impairments control devices and communicate through thought. Nevertheless, the ability of patients to move a cursor on a screen and write sentences using only their thoughts is already a reality. Clinical trials are led by the U.S. and China, with China currently ahead in speed. In March, China became the first country to approve the sale of invasive BCI medical devices. As of January, there were 21 participants in clinical trials for Neuralink in the U.S. The cumulative number of participants in clinical trials across the U.S., China, and Australia has exceeded 100.
Currently, BCI represents one-way communication, sending signals from the brain to machines. I believe that within ten years, two-way communication, where signals from machines can enter the brain, will become possible. Just as people who were skeptical after witnessing AlphaGo began to accept AI after interacting with ChatGPT, BCI will likely follow a similar trajectory. While it currently appears as a medical device for patients, the moment it transitions to use by healthy individuals, the market's size and nature will change dramatically. BCI could ultimately become a standard device replacing mobile phones and computers, raising questions about whether humans can maintain control over AI. However, in South Korea, it is still challenging to find actual patient procedures or substantial clinical research in this field. South Korea is just beginning to recognize the existence of this area. How well South Korea can catch up will determine whether it can maintain its position as a top-three technological nation. Our past experiences in semiconductor and shipbuilding industries, where we rose to global prominence, serve as a reminder. We must send doctors and engineers to the U.S. and China for training and expedite national-level budgets and systems. There are also numerous regulatory challenges to address, such as the clinical approval process for invasive devices, standards for protecting brain signal data, and long-term tracking systems post-procedure. If technology continues to advance while regulations lag behind, both patients and researchers will lose their way.
At this turning point in civilization, the role of intellectuals is to seek new directions. If we miss this direction, we will find ourselves swept along by change. Changing educational standards, setting new human goals, establishing institutions to manage risks, and investing in technologies that connect humans and machines are all tasks that cannot be postponed. A turning point, once passed, cannot be revisited.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.