ASI (Artificial Superintelligence)

Hypothetical AI that exceeds human intelligence across all domains by a wide margin — usually framed as the level beyond AGI.

ASI — Artificial Superintelligence — is the hypothetical level beyond AGI: AI that doesn't just match human ability across the board, but exceeds the best humans by a wide margin, in every cognitive domain. Where AGI is "as smart as a thoughtful adult professional", ASI is "smarter than the best human at literally everything". It matters because ASI is the focus of long-term safety arguments. If ASI arrives, even small misalignments between the system's goals and human values could cause catastrophic outcomes, because we may not be able to oversee or correct a system smarter than us. Nick Bostrom's "Superintelligence" (2014) framed much of the modern conversation; Sam Altman, Dario Amodei, and Demis Hassabis have all publicly discussed ASI timelines. The practical question: does ASI follow naturally once you have AGI (recursive self-improvement, the model trains its own successor) or does it require extra breakthroughs? Researchers disagree. "Fast takeoff" scenarios assume months between AGI and ASI; "slow takeoff" assumes years or decades. For most builders, ASI is not a near-term concern — current models are still very far from AGI in many ways. But ASI shapes how labs prioritize safety research and how governments think about AI policy. Related: AGI, alignment, recursive self-improvement, AI safety.