takeoff speed

When a plane leaves the ground, the question is not only whether it climbs but how steeply. Takeoff speed asks the same about AI: assuming systems eventually reach and pass human-level capability, how fast does that transition unfold, over years and decades, or over months, weeks, or even days?

The debate is usually framed as fast (or 'hard') takeoff versus slow (or 'soft') takeoff. In a fast-takeoff picture, often tied to recursive self-improvement, a system crosses a threshold and then improves so quickly that the world has little time to react, perhaps a near-vertical jump in capability. In a slow-takeoff picture, capability rises steeply but more continuously, spread across many competing systems and visible warning signs, giving society time to adapt, regulate, and course-correct. Note a subtlety: 'slow' here can still mean transformative change within a few years, just gradual enough to see coming, not centuries.

Takeoff speed matters because it changes which safety strategies even make sense. If takeoff is slow and continuous, we may learn from smaller failures, iterate on alignment, and build governance as we go. If it is fast and discontinuous, we might get one critical try with little warning, raising the stakes on solving alignment and control in advance. Crucially, this is unresolved: thoughtful researchers line up on both sides, the evidence is indirect, and how fast real systems would take off remains genuinely open.