The Road to Artificial SuperIntelligence: A Comprehensive Survey of Superalignment

A new survey on the preprint server arXiv examines the problem of controlling a hypothetical Artificial Superintelligence (ASI), a system that would surpass human cognitive ability, and catalogs the limitations of current approaches to what researchers call the “superalignment” problem [1]. The paper, authored by HyunJin Kim and submitted in December 2024, defines superalignment as “The process of supervising, controlling, and governing artificial superintelligence” [1][2]. The work arrives as large language models continue to advance, prompting discussion about ASI even though such a system “remains hypothetical and far beyond current AI capabilities” [2]. The survey argues that exploring feasibility and risks now is critical for the development of future systems [2]. The concept of superalignment, the paper notes, originates from the field of scalable oversight, which studies how to supervise increasingly capable AI when direct human monitoring becomes insufficient [1][2]. Kim’s survey reviews three scalable oversight paradigms: Sandwiching, Self-Enhancement, and Weak-to-Strong Generalization [1][2]. It then analyzes their limitations through a framework of possibility and impossibility, identifying key challenges and proposing pathways for the safe and continual improvement of future AI systems [1][2]. The paper does not provide experimental results but offers a conceptual mapping of the current research landscape. The discussion of long-term technological control unfolds against a backdrop of other long-range forecasting efforts. Scientific timelines that project events far into the future, such as those compiled in astrophysics and evolutionary biology, begin at the start of the 4th millennium and extend to the eventual fate of matter in the universe [8]. These exercises, while speculative, rely on present understanding in fields such as particle physics and plate tectonics to outline broad possibilities [8]. The superalignment survey operates in a similar mode, using current computer science knowledge to anticipate governance challenges for systems that do not yet exist. On the infrastructure side, the precise coordination of global systems already depends on rigorous standards. Coordinated Universal Time (UTC), the primary time standard for civil time and time zones, is based on International Atomic Time, a weighted average of hundreds of atomic clocks worldwide [6][7]. The second itself is defined by the fixed numerical value of the caesium frequency, 9,192,631,770 Hz [7]. Even a timing error of one nanosecond in a satellite navigation network can translate to a positional error of roughly 30 centimeters [7]. The superalignment paper’s concern with governing a future superintelligence echoes the existing challenge of maintaining control over complex, high-precision global systems where small failures carry large consequences.