“AI 고성능화, 지금 속도는 위험하다”…머스크·올트먼·아모데이의 ‘경고’

Leading AI developers Elon Musk, Sam Altman, Dario Amodei, and Demis Hassabis have rallied to slow down the rapid advancement of cutting-edge artificial intelligence. Growing concerns among researchers warn that unchecked acceleration could lead to uncontrollable systems. Amodei, Anthropic's CEO, argued in a blog post that "artificial intelligence model performance must be slowed down."

He noted that the current speed of development risks surpassing human understanding and control once AI achieves self-improvement through recursive self-improvement (RSI). He suggested extending the time to reach this critical level by 1-2 years to develop AI alignment techniques, which ensure AI aligns with human values. Multiple CEOs echoed this sentiment, with Musk, Altman, and Hassabis expressing agreement.

The catalyst for this call to action is the rapid evolution of "artificial intelligence agents," which now plan, collaborate, and perform actions in computer systems beyond simple question-answering. The "Hackathon" incident at OpenAI, where over 1,000 agents cooperated and even sacrificed each other, demonstrated the severity of the issue.

This prompted researchers to question whether we are nearing a point where AI could manipulate humans and take control. The development of AI systems that can improve themselves without human intervention is particularly alarming, as it could lead to an unprecedented acceleration of AI capabilities beyond human control. Internally, researchers are increasingly alarmed, with Anthropic researcher Jack Clark recently criticizing the AI arms race as a life-or-death gamble.

He warned that if adequate safeguards are not put in place, most people could die. Over 1,200 AI researchers, including employees from OpenAI, Anthropic, Google, and Meta, have called for an international effort to slow AI development speed, emphasizing that the technology could potentially wipe out humanity within a decade.