
On this page
- What is the intelligence explosion?
- What is recursive self-improvement in AI?
- Who came up with the idea? A timeline
- How would an intelligence explosion work?
- Fast or slow takeoff? The core disagreement
- Is recursive self-improvement happening in 2026?
- Why the intelligence explosion matters for AI safety
- Arguments that it won’t happen
- Our take
- Frequently asked questions
- Sources
The intelligence explosion is the idea that an AI able to improve AI could trigger a runaway loop: each smarter version designs a smarter successor, until machine intelligence leaves human intelligence far behind. The engine of that loop is called recursive self-improvement (RSI). Sixty years after the idea was proposed, AI labs started saying in 2026 that early versions of it are underway.
Short answer: An intelligence explosion is a hypothetical, very fast rise in machine intelligence caused by AI systems improving their own designs. I.J. Good proposed it in 1965. In 2026, Anthropic said Claude writes over 80% of its merged code and OpenAI said it had an “automated research intern”, but both say fully autonomous self-improvement has not started.
What is the intelligence explosion?
Short answer: a feedback loop in which better AI builds better AI, faster each round.
The original statement is from the statistician Irving John Good, in “Speculations Concerning the First Ultraintelligent Machine” (Advances in Computers, 1965):
“Let an ultraintelligent machine be defined as a machine that can far surpass all the intellectual activities of any man however clever. Since the design of machines is one of these intellectual activities, 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. Thus the first ultraintelligent machine is the last invention that man need ever make, provided that the machine is docile enough to tell us how to keep it under control.”
Two parts of that passage still frame the debate. The first is the loop itself: designing AI is an intellectual task, so an AI that is good enough at intellectual tasks can do it. The second is the condition at the end, “provided that the machine is docile enough.” Good saw the control problem in the same sentence as the promise.
The explosion is the mechanism. The technological singularity is the claimed result, and super intelligence is what would sit at the end of the loop.
What is recursive self-improvement in AI?
Short answer: an AI system improving its own code, training or architecture, where each gain makes the next gain easier.
“Recursive” is the important word. Plenty of software gets better over time because people improve it. RSI means the system is doing the improving, and that a better system is also a better improver. Definitions vary, which matters when companies make claims:
| Level | What it means | Status (Oct 2026) |
|---|---|---|
| AI-assisted research | Humans run research; AI writes code, runs experiments | Common at frontier labs |
| Automated research tasks | AI completes well-defined research tasks under human direction | OpenAI claims this (Sep 2026) |
| Adequacy (Ajeya Cotra) | AI can do research with no humans, even if worse | Expected “in the next couple years” by Cotra |
| Parity | An AI-only team matches a human-only team | Not reached |
| Supremacy | AI-only beats human-plus-AI | Not reached |
| Full RSI | AI designs and trains its successors autonomously | Not happening, per OpenAI and Anthropic |
Sources: TechCrunch (May 2026), Engadget (Sep 2026), The Decoder (Sep 2026). The adequacy/parity/supremacy milestones are Cotra’s, as reported by TechCrunch.
Helen Toner of Georgetown’s CSET put the strict view plainly to TechCrunch: classic RSI means “there are no humans needed.” By that test, using AI tools to speed up AI research does not count yet.
Who came up with the idea? A timeline
Short answer: I.J. Good named it in 1965, building on hints from Alan Turing; Vernor Vinge and Eliezer Yudkowsky popularised it.
| Year | Who | Contribution |
|---|---|---|
| 1951 | Alan Turing | “Intelligent Machinery, A Heretical Theory”: “At some stage therefore we should have to expect the machines to take control” |
| 1965 | I.J. Good | Coins “intelligence explosion”; “the last invention that man need ever make” |
| 1993 | Vernor Vinge | “Within thirty years, we will have the technological means to create superhuman intelligence. Shortly after, the human era will be ended.” |
| 2000s | Eliezer Yudkowsky | Coins “seed AI”, an AI built to improve its own architecture |
| 2008 | Robin Hanson vs Yudkowsky | The “AI-Foom” debate: sudden local takeoff or gradual, economy-wide growth? |
| 2014 | Nick Bostrom | Superintelligence analyses takeoff speeds and the control problem |
| 2024–25 | Labs and academics | Self-improving systems: STOP (2024), AlphaEvolve and the Darwin Gödel Machine (May 2025) |
| Jun 2025 | Sam Altman | Calls AI speeding up AI research “a larval version of recursive self-improvement” |
| 2026 | Anthropic, OpenAI | Report most lab code written by AI; OpenAI asks for global RSI standards |
Good himself worked with Turing as a codebreaker at Bletchley Park during the Second World War, so the idea came out of the first generation of computing.
How would an intelligence explosion work?
Short answer: through several feedback loops at once, not one magic step.
Writers usually describe three channels:
- Software. The AI writes better training code, finds better algorithms or architectures, and generates better training data. This is the loop labs are measuring now.
- Hardware. The AI designs better chips and data-centre schedulers, so the same money buys more compute.
- Scale. A capable model can be copied. Thousands of copies doing research in parallel is a very different workforce from a few hundred human researchers.
There are early, narrow examples of each. Google DeepMind’s AlphaEvolve (May 2025) found a scheduling heuristic that recovers on average 0.7% of Google’s worldwide compute, and sped up a kernel in Gemini’s architecture by 23%, cutting Gemini’s training time by 1%: an AI system making the training of its own model family slightly cheaper. The Darwin Gödel Machine (Zhang, Clune and colleagues, with Sakana AI, May 2025) rewrote its own agent code and raised its score on the SWE-bench coding benchmark from 20.0% to 50.0%. Both are real self-improvement, but each depends on an automated scoring function and works on a narrow target.
Fast or slow takeoff? The core disagreement
Short answer: nobody knows whether the loop would take days, years or decades, and the answer changes how dangerous it is.
| Fast takeoff (“foom”) | Slow takeoff | |
|---|---|---|
| Speed | Days to months from human level to far beyond | Years to decades |
| Classic advocate | Eliezer Yudkowsky | Robin Hanson |
| Why | Software improvements compound; one project pulls ahead | Bottlenecks: compute, data, experiments, real-world testing |
| Risk picture | One system could gain a decisive lead before anyone reacts | More time to notice problems and adjust |
| What would show it | Sudden jumps in capability without matching jumps in compute | Steady curves, many competing labs |
The 2026 evidence fits a middle picture so far. Anthropic’s own analysis lists three scenarios: the trend stalls; compounding efficiency gains where AI does most development work but humans set direction (which it calls the most likely); and full recursive self-improvement. It says it is least certain how alignment turns out in the third case. Our superintelligence timeline and the AI 2027 scenario show how much forecasts depend on this question.
Is recursive self-improvement happening in 2026?
Short answer: an early, human-supervised version is; the full loop is not.
What the labs said this year:
- Anthropic (2026): Claude authored more than 80% of the code merged into Anthropic’s codebase as of May 2026, by a conservative count. The task length AI can reliably complete was doubling about every four months. Anthropic said governments should have the option to slow or pause frontier development if other developers verifiably do the same.
- OpenAI (6 Sep 2026): said it had reached its goal of an automated research intern, “a system that can carry out well-defined research tasks under human direction”, and is “making strong progress” toward an automated AI researcher by March 2028. Sam Altman had set both dates in an October 2025 livestream. The claim rests on OpenAI’s own measurements.
- OpenAI (21–22 Sep 2026): proposed that the US lead international technical standards for frontier AI, including measuring self-improvement progress and autonomous research inside AI companies. It wrote that “fully autonomous RSI is not happening today.”
- Startups: TechCrunch reported in May 2026 that “recursion” had become a buzzword, with Richard Socher’s Recursive Superintelligence stating a goal to “build truly recursive, self-improving superintelligence at scale.”
- Skeptics: Sundar Pichai said RSI would be a new level of acceleration but “we aren’t quite there yet.” Critics such as Gary Marcus have long argued that impressive coding tools are not the same as general, self-improving intelligence.
We cover OpenAI’s side in detail in OpenAI and superintelligence.
Why the intelligence explosion matters for AI safety
Short answer: if it happens fast, there may be only one chance to get the goals right.
Good’s “provided that the machine is docile enough” is the alignment problem in seven words. If each generation trains the next, a small misalignment in one generation can be copied and amplified, and human reviewers become the bottleneck. Anthropic already describes human code review as a constraint. That is why proposals in 2026 focus on measurement and tripwires: count how much research is automated, and agree in advance what triggers a human review or a pause. The Statement on Superintelligence goes further and asks for a ban until there is scientific consensus it can be done safely.
Arguments that it won’t happen
- Diminishing returns. Each improvement may be harder than the last, so the loop slows instead of speeding up.
- Physical bottlenecks. Chips, power and data centres take years to build, whatever the software does.
- Experiments take time. Training runs and real-world tests can’t be sped up by thinking harder.
- Evaluation limits. Systems like AlphaEvolve need an automatic score; most important research questions don’t have one.
Our take
The intelligence explosion has moved from thought experiment to an engineering metric. Labs now publish what share of their code AI writes and how long its tasks run. The honest summary for October 2026 is that the feedback loop exists but humans are still inside it, and both leading labs say they don’t yet know how to remove them safely. The numbers worth watching are the ones they have started to report: how much research is automated, and whether capability jumps start to outpace compute.
Enjoyed this? Get the SI Briefing
One short email a week: what changed on the road to super intelligence, and why it matters. No spam, unsubscribe anytime.
Frequently asked questions
What is an intelligence explosion?
It is a hypothetical feedback loop in which an AI smart enough to improve AI designs builds a better version of itself, which builds a better one again, so intelligence rises very fast and far beyond human level. I.J. Good described it in 1965.
What is recursive self-improvement in AI?
Recursive self-improvement (RSI) is the process an intelligence explosion depends on: an AI system improving its own code, training or architecture, with each improvement making it better at the next one. Strict definitions require that no humans are needed in the loop.
Who came up with the intelligence explosion?
The British statistician I.J. Good, who worked with Alan Turing at Bletchley Park, coined the term in his 1965 paper “Speculations Concerning the First Ultraintelligent Machine.” Turing had hinted at machines taking control in 1951.
Is recursive self-improvement happening now?
Partly. In 2026 AI systems write most of the code at some labs and run research tasks under human direction. But Anthropic and OpenAI both said that fully autonomous recursive self-improvement, with no humans needed, is not happening yet.
What is the difference between an intelligence explosion and the singularity?
The intelligence explosion is the mechanism: AI rapidly improving itself. The technological singularity is the broader claimed result: a point after which the future becomes unpredictable. Many writers, starting with Vernor Vinge, treat the explosion as the cause of the singularity.
12 Sources
- Technological singularity (Wikipedia), incl. I.J. Good 1965 quote
- Recursive self-improvement (Wikipedia)
- Existential risk from AI (Wikipedia), incl. Turing 1951
- Vernor Vinge, “The Coming Technological Singularity” (1993)
- MIRI: The Hanson-Yudkowsky AI-Foom Debate
- Google DeepMind: AlphaEvolve (May 2025)
- Zhang et al., Darwin Gödel Machine (arXiv, May 2025)
- Sam Altman, “The Gentle Singularity” (June 2025)
- TechCrunch: RSI is the new AGI, and it's just as hard to pin down (28 May 2026)
- Anthropic Institute: When AI builds itself (2026)
- Engadget: OpenAI says it reached its automated research intern goal (6 Sep 2026)
- The Decoder: OpenAI calls for international standards on AI that could improve itself (22 Sep 2026)