What is an intelligence explosion?

The idea is often traced back to a 1965 paper (1966 in some listings) by the British mathematician I. J. Good. He defined an “ultraintelligent machine” as one that can far surpass all the intellectual activities of any person, however clever. Since designing machines is one of those activities, he wrote, such a machine could design even better machines; there would then unquestionably be an “intelligence explosion,” and human intelligence would be left far behind.

Good went on: “the first ultraintelligent machine is the last invention that man need ever make,” but only “provided that the machine is docile enough to tell us how to keep it under control.”

Today’s researchers think of the “explosion” in terms of speed. A September 2026 paper whose authors include the AI researchers Geoffrey Hinton and Yoshua Bengio describes an intelligence explosion as one “where years of AI progress are compressed into months or less.” In an August 2026 paper, the philosopher Toby Ord focused on the time it takes to go around the loop of AI doing research and the next AI being made. For growth to shoot up without limit toward a certain point in time, he argues, that lap time must rapidly approach zero. Even without that, he adds, growth can still be faster than steady doubling.

“Singularity” (the technological singularity) was popularized by the mathematician Vernor Vinge of San Diego State University in an essay he presented in 1993 at a symposium sponsored by NASA and others. Vinge expected a huge change to come from technology creating entities with greater than human intelligence. He called this the Singularity, “a point where our models must be discarded and a new reality rules.”

The inventor Ray Kurzweil uses the same word in a somewhat different way. The publisher’s description of his 2024 book says that “by 2045, we will be able to connect our brains directly with AI, enhancing our intelligence a millionfold. This is the Singularity.”

“Superintelligence,” as the philosopher Nick Bostrom described it in 1998, is “an intellect that is much smarter than the best human brains in practically every field, including scientific creativity, general wisdom and social skills.” Put together: the loop of AI building the next AI is recursive self-improvement (the mechanism), the sudden speed-up that results is the intelligence explosion (the event), and where it may lead is superintelligence. Because “singularity” means slightly different things to different people, it helps to check whose meaning is being used.

Views vary. In 1993, Vinge wrote that the change might come “perhaps in the blink of an eye.” On the other hand, in a June 2025 blog post, Sam Altman, the head of OpenAI (the company that makes ChatGPT), wrote that “the takeoff has started,” but also that “the singularity happens bit by bit” (this is a company head’s view).

In 2025, researchers at the research group Forethought sorted intelligence explosions into three types by what AI improves: a “software” type, where AI improves AI software (such as training methods and data); an “AI technology” type, where it also improves the chips themselves (the parts that do the computing); and a “full stack” type, where it also adds chip factories. One lap takes about 3 months to train a new AI, many months to make new chips, and years to build factories, so they see the software type as the one that could happen most suddenly.

There are also numerical estimates. In 2025, Tom Davidson and Tom Houlden of the same group estimated that once AI can fully automate AI research, a software intelligence explosion has about a 60% chance of compressing more than 3 years of AI progress into less than a year, and about a 20% chance of compressing more than 10 years into less than a year. They add that the analysis “is necessarily speculative.”

The scenario (a forecast told as a story) “AI 2027,” published in April 2025, imagines an AI that is a superhuman coder in March 2027 and superintelligence by December of that year. Its authors put their median for the time from a superhuman coder to superintelligence at about a year, with wide error margins. But in November 2025 they added: “we don’t know exactly when AGI will be built,” and “2027 was our modal (most likely) year at the time of publication, our medians were somewhat longer.”

First, AI research is itself thinking work, such as writing code and designing experiments. In 1998, the philosopher Bostrom wrote that once AI reaches human level, there will be a positive feedback loop: “AIs would help constructing better AIs, which in turn would help building better AIs, and so forth.”

Second, AI can be copied. The September 2026 paper says that today only thousands of researchers work on frontier AI research, but because AI systems can be copied and run in parallel, automating this work could add the equivalent of millions more. Davidson and Houlden’s estimate also starts from a world where AI can replace each researcher at an AI company with 30 equally capable AI systems, each thinking 30 times faster than a person.

Third, software can improve without waiting for new chips. In 2025, Daniel Eth and Tom Davidson of Forethought wrote that once AI can do the work of AI researchers, it “seems at least decently likely” that software improvements alone would cause an intelligence explosion even without more computing power, though they “can’t be confident either way.” In 2025, William MacAskill and a co-author from the same group wrote that AI that can speed up research “could drive a century of technological progress over just a few years.”

In a 2017 essay, the AI researcher François Chollet argued that an intelligence explosion is unlikely. Intelligence, he wrote, works only together with the body (senses and movement) and the surroundings, so improving the brain alone cannot make it grow without end. Even in a system that keeps improving itself, other parts start to act as bottlenecks and gains shrink, so in practice such systems “tend to display linear or sigmoidal improvement.” He also wrote that scientific progress is measurably linear, and that physics made no greater progress over 1950–2000 than over 1900–1950.

Other limits are named too: the limited amount of computing power (Eth and Davidson also weigh the objection that without more computing power, software gains might dry up); the long time it takes to learn whether an improvement really worked (International AI Safety Report 2026); and improvements becoming harder to find.

In 2024, researchers at the institute Epoch AI estimated how much software gets better when more research effort is put in (the returns to research). If each doubling of software ability needs less than twice the effort of the last (returns above 1), progress keeps speeding up; if it needs more than twice (below 1), progress slowly decelerates. For chess software the estimate was about 0.83, just shy of the threshold of 1. In four other areas, such as image recognition, the central estimates were above 1, but the researchers do not believe the current data provide strong evidence that software alone could explode.

AI now helps a lot with the work of building AI. The International AI Safety Report 2026 (February 2026, by over 100 experts, backed by over 30 countries and international organizations) says AI is increasingly used to speed up AI research itself. But it also says there is minimal empirical understanding of the feedback loops from AI automating its own research and development.

Anthropic (the company that makes the AI Claude) reported that as of August 2026, Claude “leads” 26% of its AI research and development work (a person gives a high-level prompt and supervises), up from under 1% in February 2026. It also says Claude is not operating fully autonomously for any measured part of that work (these are the company’s own measurements).

In a 2026 report, Anthropic also wrote, “We are not there yet, and recursive self-improvement is not inevitable.” Every capability it could measure has so far followed the same curve, it says, and “we have not yet seen that curve bend” (this is the company’s view).

The September 2026 paper by Hinton, Bengio and others says an intelligence explosion could be the most consequential technological development in history, and urges policymakers, including heads of government, to understand and prepare for it now. It also says some experts think AI research and development could be fully automated within the next few years.

If change comes too fast, society may not have time to prepare. The International AI Safety Report 2026 says it is still unclear how much automating AI research will speed up AI progress, and that this uncertainty creates extreme planning challenges for policymakers and institutions: if each AI advance that speeds up AI research also helps the next one, “decades of progress could happen in years.”

The September 2026 paper says an intelligence explosion could bring unprecedented breakthroughs in science and industry sooner, but also risks arriving faster than society can adapt (such as AI-enabled pandemics, cyber attacks and large-scale job disruption), and the risk that people have little oversight of automated AI research and superintelligent systems escape human control. So it proposes steps such as letting governments see how far AI research at companies has been automated, and setting a “speed limit” on how fast capabilities grow.

Back in 1965, Good added the condition that the machine be “docile.” How to keep AI safe has been tied to the idea of an intelligence explosion from the very start.