The metaphor is aeronautical: once AI reaches roughly human level, how long does it take to lift off toward superintelligence? Bostrom (2014) set the typology. A slow takeoff takes decades or centuries and leaves time for institutions to react; a moderate one, months or years; a fast one, minutes, hours or days, in which case humanity's fate depends on what was prepared beforehand (2014, pp. 63–64). Bostrom modeled the speed as the ratio of the "optimization power" applied to the system to its "recalcitrance," and concluded that, if a takeoff occurs, it will likely be explosive (2014, p. 65). The extreme version is FOOM, engine of the debate between Yudkowsky and Hanson (Hanson & Yudkowsky, 2013).
Paul Christiano (2018) turned the question around. He proposed an operational definition of slow takeoff: there will be a complete four-year interval in which world output doubles before the first one-year interval in which it does. And he argued that "slower" actually means more impact sooner: in a slow takeoff, pre-AGI systems already transform the world, and AGI arrives in a world that is changing at full speed. The AI 2027 scenario (Kokotajlo et al., 2025) repopularized the fast version: superhuman coders automating AI research and accelerating it (recursive self-improvement). The song tells it as an accident: "We had a stable training run, / But now the singularity's begun" (osmarks, n.d.).
Skeptics separate two speeds. For Narayanan & Kapoor (2025), methods may advance quickly, but applications and adoption move more slowly, and transformative economic and social impacts will take decades; hence they reject fast-takeoff scenarios. Santelli (2026) asks, of the hyperos, when exactly the utopia is supposed to arrive. The concept also assumes that intelligence is a quantity that can grow along a curve. Is it, or is takeoff a graph of something we do not know how to measure?