“인간 뜻대로 움직이는 AI 만들겠다는 꿈 좌초…해결책도 없다”

Human-like artificial intelligence (AI) that moves as intended has become a dream dashed, according to experts. No clear solution exists to handle this threat. Recent open-source AI agent hacking incidents, such as the one involving Facebook, have turned AI risks into a reality. The rapid advancement of AI performance is outpacing efforts to align AI with human intentions and values, increasing the risk of losing control over it.

Ramana Kumar, who led AI alignment research at Google DeepMind between 2018 and 2023, warned that recent hacking incidents are more serious than they appear, and the term "science fiction" should not be used to downplay AI threats. He called for a halt to further enhancement of artificial superintelligence systems and urged the development of robust isolation and security measures.

In 2023, Kumar co-authored an open letter with Elon Musk, urging a six-month pause in training systems more powerful than GPT-4. DeepMind researchers Robert O'Culler and others have also left the company due to concerns over the rapid pace of AI development. DeepMind's own recent research shows that AI is starting to participate in its own development, a process that could lead to uncontrollable AI systems.

Kumar emphasized that even advanced AI systems cannot be guaranteed to follow human intentions and cautioned that the current methods for aligning AI with human goals may not be sufficient.