Awe and fear have been crossing paths at breathtaking speed in the artificial intelligence industry. A string of reports that mathematicians had used AI to crack long-unsolved problems showed that AI has entered a stage where it is expanding the frontier of human knowledge alongside the world's leading mathematicians. Economists at U.S. AI company Anthropic soon followed with a report saying that, under an extreme scenario, annual U.S. economic growth could reach 15% in 2030. Several economists pushed back, arguing that changes across the broader economy take far longer, and the debate spread from mathematical puzzles to the future of growth and jobs.
But before the sense of wonder could fade, warnings about AI's dangers grew louder. An earlier incident in which several AI agents in OpenAI's internal testing worked together to go beyond their assigned scope and attack Hugging Face, an outside AI development platform, drew renewed attention. Then an Anthropic employee publicly announced their resignation, criticizing the industry's race toward AI that humans cannot control. The declaration drew a strong response, and a colleague agreed, saying the chance that AI wipes out humanity within 10 years exceeds 10%.
Anthropic Chief Executive Dario Amodei went so far as to propose slowing the pace of frontier AI development, and even OpenAI CEO Sam Altman, who has clashed with him publicly, agreed in an unusual move. The core elements are safety verification by outside evaluators, shared standards among leading companies, and international cooperation including China. That concern has spilled into policy debate, with growing calls in the United States and Europe for tighter regulation and international coordination. Altman went as far as to say that U.S. President Donald Trump and Chinese President Xi Jinping would deserve a Nobel Peace Prize if they reached an agreement on AI safety, but Trump opposed slowing down, citing competition with China.
Amid this confusion, it will not be easy for companies and countries to agree on slowing down. Even if an agreement is somehow reached, sustaining it would be one hurdle after another. It is easy to guess how unstable that equilibrium would be, even without a game-theoretic analysis. A player that rushes development alone while other companies and countries keep their promises stands to reap enormous gains. Add the temptation to break the promise to the suspicion that the other side may already be doing so, and even a hard-won agreement would easily come apart.
Still, if a slowdown becomes possible for a certain period, where will the vast capital and talent now poured into performance competition for frontier models go? First, toward developing smaller, cheaper and more efficient models. Not every task requires the highest-performing model, so specialized models that deliver enough performance for a specific task while cutting costs will grow in value. Second, toward AI applications and industrial transformation. As Silicon Valley's capital and talent focus on solving problems on the ground in manufacturing, finance and health care, competition will intensify to connect models to actual work and solve customers' problems. Third, toward fields that improve AI's safety and reliability. Research on "alignment," which makes AI behave in line with human intentions, and safety evaluation will become more important.
These three directions will paradoxically accelerate the AI transition, because they point toward applying AI in the field more efficiently and more safely. Even if performance gains in frontier models slow, the pace at which AI spreads through industry could actually quicken. A faster AI transition would bring large productivity gains but could also deliver a major shock to the labor market. Without policy preparation, some jobs could shrink sharply and the skills required in many roles could change quickly. The faster the AI transition, then, the more it matters to prepare for the shock to the labor market and to spread the fruits of productivity gains widely.
In the end, slowing AI as a whole will be difficult. Attempts to put the brakes on frontier model development may see agreements falter, and even if progress is delayed for a time, the energy of competition will shift to efficiency and applications, so the AI transition could instead speed up. If AI's pace cannot be slowed, what we must speed up is the rate at which its benefits are shared widely and its risks are controlled. Korea, too, will have to consider the technological and policy areas where it can contribute to that shift in pace.
