OpenAI’s Ilya Sutskever Suggests Building a Bunker for Researchers
Ilya Sutskever, a co-founder of OpenAI, once proposed constructing a bunker intended to safeguard the company’s leading researchers in case of a catastrophic event, described as “Rapture,” stemming from the emergence of artificial intelligence (AI).
This intriguing detail emerges from Karen Hao’s forthcoming book, The Empire of AI: Sam Altman’s OpenAI Dreams and Nightmares. Sutskever, often viewed as the genius behind ChatGPT, made his comments during a meeting in 2023 with prominent scientists within the organization.
During the discussion, Sutskever’s remark, “Once we all enter the bunker…” caught the attention of puzzled colleagues who responded with, “Sorry, bunker?” He clarified, stating, “We definitely plan to build a bunker before we release AGI.”
Moreover, excerpts from the book indicate this wasn’t a passing thought; sources told Hao that Sutskever frequently referenced the bunker in various internal conversations.
Concerns About Society’s Readiness for AGI
This isn’t an isolated concern coming from leaders in the AI field. Demis Hassabis, the CEO of Google DeepMind, has also expressed worries about society’s preparedness for AGI.
“I think we’re on the edge. Maybe we’re five or ten years from now. Some people are shorter, and I wouldn’t be surprised,” Hassabis mentioned, reflecting on the timeline for AGI’s arrival. He added, “It’s a sort of probability distribution… But anyway, it can come soon, but I don’t know if society is still very prepared for it.” He emphasized the importance of discussing the controllability of these systems.
Additionally, the 2024 Chemistry Nobel Prize recipient previously advocated for the establishment of overseeing bodies, resembling the United Nations, to manage the development of AGI.
Understanding AGI
AGI stands for artificial general intelligence, and it represents a significant leap from conventional AI. While AI is typically designed for specific tasks, AGI aspires to possess a form of intelligence capable of functioning across a broader spectrum of activities, similar to human cognition. In essence, AGIs would become machines that can understand, learn, and apply knowledge across various domains.
