What is AI takeoff? Fast and slow scenarios

What takeoff means

Forecasts about AI usually ask when: when might AGI arrive, when might superintelligence follow? Those are covered on our timelines page. Takeoff is a different question: once AI reaches roughly human level across the board, what does the curve look like afterwards?

The term comes from discussions in the 2000s among researchers concerned about AI risk, and was given its standard framing in Nick Bostrom’s 2014 book Superintelligence, which distinguished three speeds:

Slow

Decades or centuries. Time for institutions to adapt, laws to be passed and new problems to be studied as they appear.

Moderate

Months or years. Some time to react, but not enough for ordinary political processes.

Fast

Minutes, hours or days. Effectively no time to respond once it starts.

Bostrom judged a fast or moderate takeoff more likely than a slow one. That judgement, not the definitions, is what has been argued over since.

Fast, or hard, takeoff

The case for a fast takeoff rests on a feedback loop. If an AI system becomes good enough at AI research to improve itself, or to build a better successor, each improvement makes the next one easier, and the process accelerates. That loop is the intelligence explosion, and the mechanism behind it is recursive self-improvement.

Eliezer Yudkowsky, the most prominent advocate of the fast view, argued in a 2013 paper that the key question is the “returns on cognitive reinvestment”: whether putting more intelligence into improving intelligence yields more than proportionally better results. If it does, he argued, progress could jump from roughly human-level to far beyond it very quickly, possibly within a single project, and the first system to get there could gain a decisive lead.

Other reasons given for expecting speed:

• Copying. A human expert takes decades to train. A trained AI system can be copied as many times as there are computers to run it on.
• Hardware overhang. If the software breakthrough comes after the chips already exist, the new system may immediately be able to run at enormous scale.
• Human level is not special. There is no reason to expect machine capability to slow down as it passes the level of the people it is measured against.

Slow, or soft, takeoff

The main alternative was set out by Paul Christiano, an alignment researcher then at OpenAI, in a 2018 essay on his personal blog, “Takeoff speeds”. He did not argue that change would be gentle. He argued that it would be continuous: that the world would see increasingly powerful AI transforming the economy before any single system crossed a dramatic threshold. He gave a test: a slow takeoff is one in which there is a complete four-year period in which world economic output doubles before the first one-year period in which it does. That would itself be far faster growth than anything in history.

The reasons offered for expecting a more gradual curve:

• Bottlenecks. Progress needs computing power, chips, electricity and experiments that take real time, none of which a smarter system can produce instantly.
• Competition. Many organisations are working at similar levels, so a slightly better system replaces a slightly worse one rather than leaping far ahead.
• Diminishing returns. Each improvement may get harder to find.
• Precedent. Most technologies, including AI so far, have improved along fairly smooth curves.

The two positions were argued out at length in a 2008 online debate between Yudkowsky and the economist Robin Hanson, later published as The Hanson-Yudkowsky AI-Foom Debate (“foom” being the fast scenario). It remains a useful record of the arguments on both sides.

Continuous versus discontinuous

Speed and smoothness are different things, and the clearest writers keep them apart. A takeoff can be fast but continuous, with progress accelerating sharply along a smooth curve, or slow but discontinuous, with a long plateau followed by a sudden jump. The practical question underneath both is whether the world would get warning: weaker versions of the dangerous capabilities, appearing early enough to study and prepare for.

The evidence from the last few years can be read both ways. Many measures of AI capability have improved smoothly and predictably as computing power has grown; METR’s measure of how long a task AI can complete on its own has doubled at a fairly regular pace since 2019. But the same measure has sped up since 2024, and individual skills sometimes appear abruptly when measured on a pass-or-fail basis.

Is there a human-level threshold at all?

Takeoff is usually described as what happens after AI reaches “human level”. That assumes a threshold exists. It may not. Today’s systems are already far better than people at some tasks and worse at others; the International AI Safety Report 2026 describes them as excelling at some difficult tasks while failing at simpler ones. Rather than crossing a line, AI may keep becoming superhuman at more things one at a time. Our page on recognising AGI explains why there is no agreed test for the threshold itself.

The threshold that may matter more is narrower: the point at which AI can do most of the work of AI research. That is where the fast-takeoff argument says acceleration would begin. A 2023 model by Tom Davidson, then at Open Philanthropy, estimated that the time from AI being able to do 20% of cognitive work to being able to do all of it would be around three years, faster than most economists would expect but slower than the fastest scenarios. The 28 September 2026 working paper on the intelligence explosion cites tentative extrapolations suggesting that months-long AI research projects could be automated by mid-2028; that is a forecast, not an observation.

Why the shape matters

The speed of takeoff determines what kind of safety strategy makes sense. In a slow takeoff, problems can be found in weaker systems and fixed, institutions can adapt, and governments can regulate as they go. In a fast one, the first attempt may be the only one, which is why researchers who expect speed put so much weight on solving alignment and control in advance. Much of the current policy debate, including calls in September 2026 for governments to monitor how far companies have automated their own research, is an attempt to find out which kind of takeoff, if any, the world is in, before it is too late to matter.