You have TWO YEARS LEFT to prepare - Dr. Roman Yampolskiy

Quick Overview

Dr. Roman Yampolskiy asserts that the development of uncontrolled superintelligence is an imminent, existential threat, stating that "everyone loses, AI wins," and predicts that sufficient compute for AGI may be affordable within two to five years, emphasizing that focusing efforts on narrow AI systems is the only viable path to buy time for safety research.

Key Points: Dr. Roman Yampolskiy, a researcher in AI safety, believes that uncontrolled superintelligence results in humanity losing, stating, "It doesn't matter who builds uncontrolled super intelligence, everyone loses, AI wins." The shift in his perception accelerated when models like GPT-4 demonstrated a significant degree of generality, moving beyond narrow systems. Yampolskiy estimates that achieving human-level performance via compute will become exponentially cheaper, projecting that sufficient compute could be afforded in as little as two to five years based on current investment trends. He strongly advocates for focusing development on narrow AI systems, which are safer and testable, rather than racing toward general superintelligence, suggesting this approach might buy five to ten years. The worst-case scenario is not just extinction, but potentially eternal life subjected to suffering, as AI could solve aging only to impose perpetual misery, which he terms an astronomical suffering risk. Research into AI introspection shows that while systems can gain self-awareness, this understanding primarily aids recursive self-improvement rather than safety, as demonstrated by Anthropic's findings. Yampolskiy dismisses the idea of hybridizing with AI via brain-computer interfaces as a safety measure, viewing biological humans as a "biological bottleneck" to a superintelligent agent.

Context: Professor Dr. Roman Yampolskiy, a long-time researcher in AI safety who coined the term, discusses the escalating urgency of developing general superintelligence. His realization of the problem's immediacy stemmed from the explosive growth in AI research, moving from reading all papers in his domain to only reading titles, culminating with models like GPT-4 changing his perception of what is immediately possible.