Fast Takeoff vs Slow Takeoff in Artificial Intelligence
AI takeoff describes how quickly capabilities might grow around and beyond broadly human-level performance. It is different from a timeline forecast: the timeline asks when transformative AI arrives, while takeoff asks what happens to capability after it does.
Fast and slow takeoff are ends of a spectrum, not two guaranteed futures. The speed, concentration, and visibility of progress determine how much time institutions have to learn and respond.
The fast-takeoff case
In a fast or hard takeoff, capability improvement accelerates over days or months. A system might help improve AI research, automate experiments, copy itself, or rapidly convert resources into further capability. If gains remain concentrated, one system or organization could develop a decisive advantage before competitors and regulators adapt.
The strongest version is sometimes called an intelligence explosion or “FOOM.” It remains a contested forecast, especially because software improvement still depends on hardware, data, experiments, coordination, and diminishing returns.
The slow-takeoff case
In a slow or gradual takeoff, progress unfolds across years and many organizations. AI systems transform research and the economy before any one system becomes overwhelmingly capable. Failures may become visible earlier, giving society time to test controls and build institutions.
Gradual change is not automatically safe. Competitive pressure can spread risky systems, concentrate wealth and influence, or normalize increasing autonomy. Distributed harms may be harder to attribute than one dramatic event.
Why the distinction changes strategy
Fast takeoff puts weight on safeguards that work before deployment, secure model development, and plans for sudden capability jumps. Slow takeoff puts more weight on monitoring, standards, broad coordination, labor and economic adaptation, and learning from real incidents.
A robust approach should not bet everything on one speed. Warning indicators might include rapid automation of AI research, unusually steep benchmark gains, expanding autonomous tool use, or capability diffusion across institutions.
How simulations use takeoff assumptions
A scenario model can vary resource availability, replication speed, human coordination, and crisis frequency to represent different takeoff conditions. The result is not a forecast; it reveals which strategies and safeguards depend on having more time.