AI Takeoff Speed
AI takeoff speed is the time an AI-development trajectory takes to move between explicitly stated capability milestones. It is a comparison frame, not one forecast. Slow or soft and fast or hard overlap in usage but are not standardized pairs; FOOM is the stronger historical shorthand for an explosive scenario, commonly associated with rapid feedback or self-improvement. A useful claim states its start, endpoint, capability metric, actors, and calendar interval rather than treating these labels as interchangeable.
Origin and context
I. J. Good's 1965 intelligence-explosion argument supplied an earlier feedback-loop idea. In December 2008, Yudkowsky's Hard Takeoff essay used AI go FOOM within a debate with Robin Hanson over whether generally intelligent systems could improve very quickly. Bostrom's 2014 book later made takeoff speed part of superintelligence analysis. Contemporary models retain the question but operationalize it differently: Davidson measures the interval from systems able to automate 20% to 100% of cognitive tasks, weighted by economic value. These dates are evidence anchors, not a claim that one author coined every label.
Why it matters
Takeoff speed matters because it changes the time available to test systems, interpret warning signs, coordinate institutions, adapt work, and deploy safeguards. It does not by itself determine whether development is continuous, whether one actor leads, or whether an intelligence explosion occurs. Carlsmith separates fast, discontinuous, concentrated, feedback-driven, and recursive-self-improvement scenarios. A short transition may result from compute, algorithms, investment, deployment, or feedback; a feedback loop can also accelerate and then peter out.
Example
Suppose one study defines its start as systems that can automate 20% of cognitive tasks and its endpoint as 100%, then estimates an interval. Another asks how long it takes to move from human-level general intelligence to broad superintelligence. Even if both call their result fast takeoff, they answer different questions and cannot be compared without translating milestones. Conversely, a sudden jump on one benchmark is not by itself hard takeoff: it may be narrow or unrelated to the chosen endpoint. This page therefore treats soft versus hard as a family of scenario comparisons, not a measured binary property of current models.
How it differs
Superintelligence
Superintelligence is a capability level or destination. Takeoff speed describes the duration and dynamics of moving between levels. A slow path could still end in superintelligence, while a fast local jump does not establish that the destination has been reached.
Capability overhang
Capability overhang is a stored enabling condition, not a rate. In older discussions it often means available compute or resources awaiting adequate software; the local record also uses a newer latent-capability and elicitation sense. Either may contribute to a fast transition, but neither specifies the interval or guarantees an intelligence explosion.
AGI
AGI is a contested capability threshold; takeoff speed concerns movement between explicitly defined thresholds. Using post-AGI as a starting point without an operational test makes duration claims difficult to compare.
Maturity and evidence
Maturity is rated 3. The vocabulary has multi-decade continuity, independent analytical frameworks, and current computational models. But there is no agreed milestone pair, capability scalar, or boundary between soft or slow and hard or fast, and the relevant transition has not been empirically observed. Maturity would rise with stable operational definitions and retrospective evidence across several capability measures.
Limits and open questions
These are conditional scenarios, not measurements or forecasts endorsed by Skills Intelligence. FOOM often implies a stronger feedback-driven story than merely fast, and authors vary on whether hard means rapid, discontinuous, concentrated, or all three. Current models depend on uncertain assumptions about compute, algorithms, automation, bottlenecks, and feedback generation time. Ord's 2026 analysis is a preprint and argues that singular growth requires stronger conditions than some simpler models assume.
Related terms
References
- Hard TakeoffEliezer Yudkowsky / LessWrong · 2008-12-02 · class A
- Intelligence Explosion MicroeconomicsMachine Intelligence Research Institute · 2013-09-13 · class A
- Superintelligence: Paths, Dangers, StrategiesOxford University Press / Nick Bostrom · 2014-07-03 · class A
- Is Power-Seeking AI an Existential Risk?Joseph Carlsmith / arXiv · 2022-06-16 · class A
- What a Compute-Centric Framework Says About Takeoff SpeedsCoefficient Giving / Open Philanthropy · 2023-06-27 · class B
- An interactive model of AI takeoff speedsEpoch AI · 2023-01-24 · class B
- The Dynamics of Intelligence ExplosionsToby Ord / arXiv · 2026-08-14 · class A
Last updated: 2026-09-05
This term is also covered in the Skills Atlas as ai risk management skill.
This term is also covered in the Skills Atlas as model evaluation skill.