When will AI be created?

Abstract

This article delves into the challenges of developing artificial general intelligence (or ‘safe’ AGI), exploring various approaches and perspectives held by researchers. Five major types of approaches are described, including addressing threat models to identify risks and vulnerabilities; agendas to build safe AGI systems; robustly good approaches, which prioritize generalizability and resilience; de-confusing complex concepts and aligning motivations; and field-building, aiming to expand the research community to meet the challenges of AGI development. The discussion centers on three main threat models: Power-Seeking AI, Inner Misalignment, and AI Influenced Coordination. It then explores three proposed agendas to build safe AGI: Iterated Distillation and Amplification (IDA), AI Safety via Debate, and Solving Assistance Games. Within the ‘Robustly Good Approaches’ category, emphasis is placed on interpretability, robustness, and forecasting. Lastly, four key considerations are examined: Prosaic AI Alignment, Sharpness of Takeoff, Timelines, and the Difficulty of Alignment. – AI-generated abstract.