AI takeoff: hard versus soft

An AI takeoff is the transition from roughly human-level AI to superintelligence. A hard (or fast) takeoff happens in minutes, days or months, usually inside one project. A soft (or slow) takeoff stretches over years or decades, with AI transforming the economy along the way. The two words describe how fast and how suddenly the change happens, not whether it happens.

Where do the terms come from?#

The modern vocabulary comes from two places. Nick Bostrom’s Superintelligence (2014) sorted takeoff by duration. As Katja Grace summarized chapter 4, a “slow takeoff” takes decades or centuries, a “moderate takeoff” months or years, and a “fast takeoff” minutes to days.

“Hard takeoff” appears in Eliezer Yudkowsky’s writing by 2001. “FOOM,” an informal name for it, was used in his 2008 online debate with the economist Robin Hanson, in which Yudkowsky argued for a fast takeoff by a single system and Hanson expected growth spread across many actors. The intelligence explosion guide covers that debate.

Bostrom also gave a rough formula for the speed: the rate of change in intelligence equals optimization power divided by “recalcitrance,” which Grace glosses as “how hard it is to make the system smarter by applying effort.” Takeoff is fast if the effort applied grows quickly, especially once most of it comes from the AI itself, and the problem does not get harder at the same pace.

What is the difference between fast and discontinuous?#

Two separate questions often get merged into one. One is duration: how long the transition takes. The other is shape: whether progress follows its existing trend or jumps.

The researcher Paul Christiano gave “slow takeoff” a concrete test in 2018:

“There will be a complete 4 year interval in which world output doubles, before the first 1 year interval in which world output doubles.”

— Paul Christiano, “Takeoff speeds,” 2018

In his framing, a fast takeoff is “a breakthrough within a small group,” and a slow one is “a continuous acceleration distributed across the broader economy or a large firm.” He thought slow takeoff “significantly more likely,” but slow in his sense is not calm. He titled one section “Slower takeoff means faster progress,” because a continuous takeoff means AI is already reshaping the world before it reaches human level. He also wrote that an intelligence explosion “seems very likely” to him.

History offers some evidence on jumps. AI Impacts, a research project that studies AI forecasting, looked for sudden leaps in past technologies and found ten events that produced a robust discontinuity “equivalent to more than a century at previous rates.”

How fast do current models say it would be?#

Since 2023, takeoff has moved from argument to explicit models with numbers. Each one depends on assumptions its authors set out.

The AI 2027 authors have since revised their timelines. In December 2025 their updated model put full coding automation about 3 to 5 years later than their earlier one. In August 2026 they wrote that “reality seems to be going about 70-90% as fast as AI 2027 predicted.”

The models disagree about where the speed would come from. A March 2025 paper by Daniel Eth and Tom Davidson at Forethought argued that AI improving AI software could produce dramatically more advanced systems “within months, even with fixed computing power.” Ege Erdil and Matthew Barnett, writing for Epoch AI in a post that speaks for its authors, argued that most of AI’s impact will come from broad automation across the economy rather than from AI research. The recursive self-improvement guide covers what AI companies report about AI research being automated.

Why does the speed matter?#

A slow takeoff leaves time to notice problems, correct them, and adapt laws and institutions. A fast one, especially inside a single project, could leave one developer far ahead of everyone else, with little time to test whether its system is safe. Superintelligence calls this possibility a “decisive strategic advantage.” Bostrom argued that the control problem should be solved before superintelligence is built, because it may be impossible to solve afterward.

OpenAI’s chief executive, Sam Altman, has described a takeoff already under way. In June 2025 he wrote that “the takeoff has started,” describing a gradual process rather than a sudden break. The singularity guide discusses that framing.

Other full entries: Orthogonality thesis, Instrumental convergence, Paperclip maximizer, Control problem, Superalignment. Every term is in the glossary.