AI Takeoff
AI takeoff refers to how quickly artificial intelligence capability would grow once systems approach and then pass human-level general ability: whether the transition from "roughly human" to "vastly superhuman" would unfold over years and decades — a slow or soft takeoff — or within weeks, days or less — a fast or hard takeoff, sometimes nicknamed "foom." The question matters less as a forecast than as a hinge for strategy: if change is gradual, society and developers get repeated chances to observe, test and correct systems, whereas if it is abrupt, the first sufficiently capable system might gain a decisive advantage before anyone can react.
Nick Bostrom's 2014 book Superintelligence gave the standard framing by splitting takeoffs into fast, moderate and slow, and by expressing the rate of improvement as optimization power divided by recalcitrance. Optimization power is the effort applied to improving the system (research, compute, and eventually the system's own contribution), and recalcitrance is how hard further improvement is. A fast takeoff needs low recalcitrance, high and growing optimization power, or both, and Bostrom argued that if the system itself contributes substantial optimization power, a feedback loop could develop. That feedback loop is the subject of the intelligence explosion and recursive self-improvement entries; takeoff is specifically about the speed and shape of the resulting curve.
The best-known argument about the question is the Yudkowsky–Hanson "foom debate." In 2008 on the blog Overcoming Bias, Eliezer Yudkowsky and economist Robin Hanson exchanged long posts, later collected by the Machine Intelligence Research Institute as an ebook. Yudkowsky argued that a mind able to improve its own design could cross a threshold after which progress is sudden and discontinuous, comparing it to qualitative leaps in evolutionary history. Hanson argued from economic history that major growth transitions have been distributed across many contributors and innovations rather than a single insular process, and that a lone project is unlikely to outpace the rest of the world. Neither convinced the other, and the debate remains a reference point because it exposes the actual cruxes: how lumpy innovation is, how much progress comes from a single architectural insight versus accumulated compute and data, and whether capability gains diffuse quickly between competitors.
Paul Christiano's 2018 essay "Takeoff speeds" made the question operational and economic. He defined a slow takeoff as one in which a full four-year interval of doubling in world economic output occurs before the first one-year doubling begins, and a fast takeoff as one where the one-year doubling arrives without the preceding four-year doubling. His view was that a slow takeoff, in this sense, is more likely, because earlier, weaker AI systems would already be transforming the economy and AI research before the most capable systems arrive, so there would be no sudden unobserved leap. Later modeling work, such as Tom Davidson's 2023 compute-centric framework for Open Philanthropy, tried to quantify how long the interval between systems that could automate a fifth of economic tasks and those that could automate nearly all might last, finding that the answer depends sharply on assumptions about compute, algorithmic progress and automation of research itself.
Empirical context has added evidence without settling the issue. The 2020 scaling-law results of Kaplan and colleagues at OpenAI showed smooth, predictable improvements in language-model loss with more parameters, data and compute, which is often cited in support of continuous progress, yet particular abilities have seemed to appear suddenly at certain scales, a pattern whose reality has been disputed as partly an artifact of how performance is measured. The rapid deployment of increasingly capable models from several labs through the early 2020s looks closer to the gradual, multi-actor picture than to a single decisive jump, but a gradual start does not rule out later acceleration if AI systems come to do much of AI research.
A frequent misconception is that "slow" means safe or distant. Christiano's slow takeoff still involves explosive change within a few years by historical standards, and a slow path may still be risky because competitive pressure, deployment decisions and institutional lag matter as much as the shape of the curve. The opposite pitfall is treating foom as the default prediction of everyone concerned about AI risk; many safety researchers expect continuous progress and still worry about it.
Read AI takeoff as a debate about the shape of the capability curve rather than its destination: whether society will watch systems cross human level gradually and adapt, or face an abrupt jump, and what each answer implies about how much preparation must happen before the fact.