The intelligence explosion and the singularity

An intelligence explosion is the idea, proposed by I. J. Good in 1965, that an AI able to improve its own design could trigger a rapid chain of ever-smarter systems. The singularity is broader: a point after which technological change is so fast that the future cannot be predicted. Vernor Vinge tied it to superhuman intelligence in 1993.

What is an intelligence explosion?

An intelligence explosion is a hypothesized feedback loop in which AI systems design better AI systems, each round faster than the last.

The idea comes from I. J. Good’s 1965 paper, “Speculations Concerning the First Ultraintelligent Machine.” Good reasoned that designing machines is one of the intellectual activities a sufficiently capable machine could do better than people. So:

“An ultraintelligent machine could design even better machines; there would then unquestionably be an ‘intelligence explosion,’ and the intelligence of man would be left far behind.”

— I. J. Good, Advances in Computers, vol. 6, 1965

If each round of redesign also shortens the next round, the process runs away.

The modern name for the mechanism is recursive self-improvement. The system improves the very thing that does the improving.

The argument needs every step of the loop to keep paying off. If improvements get harder faster than the system gets smarter, the loop slows instead of exploding. Much of the disagreement is about that point.

Where does “the singularity” come from?

The word is borrowed from mathematics and physics, where a singularity is a point at which a quantity becomes undefined or infinite.

The first use in this sense is usually traced to 1958. In a tribute to John von Neumann, the mathematician Stanislaw Ulam recalled a conversation about accelerating technology that seemed to approach an “essential singularity” in human history, beyond which human affairs as known could not continue.

Vernor Vinge, 1993

Vernor Vinge connected the singularity to superhuman intelligence in “The Coming Technological Singularity,” written for a symposium sponsored by NASA Lewis Research Center in March 1993. His abstract opens:

“Within thirty years, we will have the technological means to create superhuman intelligence. Shortly after, the human era will be ended.”

— Vernor Vinge, 1993

Vinge’s thirty years ran out in 2023. Whether the “technological means” existed by then is itself a matter of definition; no superhuman general intelligence had been built.

Ray Kurzweil, 2001 to 2024

The inventor and author Ray Kurzweil made the singularity a popular idea. His 2001 essay “The Law of Accelerating Returns” argued that technological progress is exponential rather than linear, because each generation of tools is used to build the next.

In The Singularity Is Near (2005), he put the singularity at 2045. In The Singularity Is Nearer (Viking, June 2024), he kept both of his headline dates: AGI by 2029 and the singularity in 2045. Kurzweil’s singularity includes humans merging with AI, not only machines surpassing them.

“The gentle singularity,” 2025

In June 2025, OpenAI chief executive Sam Altman wrote that “we are past the event horizon; the takeoff has started.” His essay describes a gradual process in which AI, researchers, capital and infrastructure reinforce one another, rather than a single machine rapidly building its successor.

The change in framing drew comment. Nieman Lab asked when the singularity had become “gentle.” The classic idea is a sharp break. Altman’s version is a slope.

As a result, “singularity” now has at least three meanings. Vinge’s is a break in predictability caused by superhuman minds. Kurzweil’s is a dated merger of human and machine intelligence. Altman’s is an ongoing, gradual acceleration. A claim that “the singularity has started” means very little until the speaker says which.

What is the strongest case for an intelligence explosion?

The clearest modern statement is by the philosopher David Chalmers, in “The Singularity: A Philosophical Analysis” (2010). He restated Good’s idea as a chain:

1. There will be human-level AI before long, absent defeaters.
2. If there is human-level AI, there will soon be greater-than-human AI (AI+).
3. If there is AI+, there will soon be far-greater-than-human AI (AI++).
4. So there will be AI++ before too long, absent defeaters.

By “defeaters” Chalmers means events such as global catastrophes that would stop the process. The argument does not say an explosion is certain. It says that if the premises hold, superintelligence follows.

Supporters add practical reasons. Digital minds can be copied, run faster, and edited, as the types of superintelligence guide explains. AI companies are also aiming directly at the loop: OpenAI has said its goal is an automated AI researcher by March 2028. The argument is that if AI systems can do AI research, human researchers stop being the limit on how fast the loop turns.

How fast would it happen?

Proposed answers range from minutes to centuries, and speed is a separate question from whether an explosion happens at all.

Nick Bostrom’s Superintelligence (2014) distinguishes slow takeoff (decades or centuries), moderate takeoff (months or years) and fast takeoff (minutes to days). Only the fast case matches the image of an explosion.

The question was argued at length in 2008 by economist Robin Hanson and researcher Eliezer Yudkowsky, in what became known as the “AI-Foom Debate.” Yudkowsky argued for a fast, local takeoff, in which one system pulls far ahead. Hanson argued that growth would be spread across many actors and an entire economy, as with past technologies. The Machine Intelligence Research Institute collected the debate in a 2013 ebook.

What are the main arguments against?

The main objections are that intelligence depends on its environment, that physical experiments take time, that returns may diminish, and that the concept itself is unclear.

Intelligence depends on the environment

In “The Implausibility of Intelligence Explosion” (November 2017; first published under the title “The Impossibility of Intelligence Explosion”), AI researcher François Chollet argued that intelligence is situational. A brain is one part of a larger system that includes a body, an environment, other people and culture. Improving the brain alone does not improve the whole. He also pointed out that science already improves itself, through better tools and more researchers, yet its progress has looked roughly linear rather than explosive.

The Machine Intelligence Research Institute published a point-by-point reply that assessed each of Chollet’s arguments separately and disputed several of them.

The real world is slow

Many discoveries require experiments, manufacturing and testing that take physical time. A mind that thinks faster still has to wait for the lab results. Arvind Narayanan and Sayash Kapoor, in “AI as Normal Technology” (April 2025), argue that AI’s effects will spread over decades, slowed by safety requirements, organizational change and institutional adoption.

Returns may diminish

Each improvement may be harder to find than the last. If so, a self-improving system could level off rather than accelerate. This is the same point raised at the top of this page, and the evidence so far has not settled it.

The concept itself may be unclear

Narayanan and Kapoor call superintelligence “incoherent as usually conceptualized.” Chollet doubts that general intelligence exists in the sense the argument assumes. If intelligence is not a single quantity, “more of it” may not be a meaningful goal.

What do AI researchers think?

They are divided, and more lean against the argument than toward it.

The 2023 Expert Survey on Progress in AI, run by AI Impacts, asked researchers whether the intelligence explosion argument is broadly correct. Of 299 who answered:

• 9% said quite likely (81–100%)
• 20% said likely (61–80%)
• 24% said about even chance (41–60%)
• 24% said unlikely (21–40%)
• 23% said quite unlikely (0–20%)

Asked separately about superintelligence after human-level machine intelligence, the median respondent gave a 10% chance within two years and 60% within thirty years. The survey had a 15% response rate. The timelines guide covers what surveys can and cannot show.

The safety concerns that follow from a fast takeoff, and the public statements responding to them, are in the debate guide.