Fast Takeoff

A fast takeoff (or hard takeoff) is a scenario in which an artificial general intelligence system undergoes rapid, recursive self-improvement, crossing from human-level to superhuman-level capability in a timescale too short for human institutions to respond. The term was popularized by Nick Bostrom and Eliezer Yudkowsky, and it represents one of the two canonical scenarios for AGI development — the other being slow takeoff, in which capability growth is gradual and societies have time to adapt.

The fast takeoff hypothesis rests on three premises: that intelligence is a general-purpose capability that can be applied to its own improvement; that self-improvement produces compounding returns; and that the resulting capability threshold is sharp rather than gradual. If these premises hold, then the first system to cross the threshold acquires a decisive strategic advantage — a period during which no other actor can compete, because the gap in capability is too large to close.

The strategic implications are severe. In a fast takeoff scenario, the alignment problem must be solved before the threshold is crossed, because there is no opportunity for trial and error afterward. The system that crosses the threshold first determines the long-term future, and if its goals are misaligned, the outcome is not merely harmful but existentially catastrophic. This is the argument for AI safety as an urgent priority: not because AGI is imminent, but because if a fast takeoff occurs, preparation must be complete in advance.

Critics of the fast takeoff hypothesis argue that intelligence is not a single, fungible capability but a heterogeneous collection of skills; that self-improvement is subject to diminishing returns; and that real-world AI development involves distributed, overlapping systems rather than a single agent. The empirical trend toward slow, capability-limited improvement in current systems supports the slow takeoff view. But the fast takeoff argument does not depend on current trends. It depends on the possibility of a phase transition — a discontinuity that current data cannot rule out.

The systems-theoretic assessment is that fast takeoff is not merely a prediction about AI. It is a prediction about the dynamics of capability amplification in complex systems. Whether such amplification produces smooth or discontinuous growth depends on the architecture of the system, the structure of its environment, and the feedback loops that govern its self-modification. These are empirical questions, but they are empirical questions about a system that does not yet exist.