What is an intelligence explosion, and is it happening? (2026)
An intelligence explosion is a feedback loop where AI gets good enough at AI research to improve itself, each round faster than the last. As of October 2026 a partial, human-supervised version runs inside the big labs, but no lab claims full recursive self-improvement yet.
An intelligence explosion is a feedback loop where an AI gets good enough at AI research to improve itself, and each improvement makes the next one come faster, until it ends up far beyond human intelligence. As of October 2026 the loop has started in a partial, human-supervised form inside the big labs. The runaway version has not happened.
That's the short answer. The longer one matters if you build AI products, because the labs that sell you your model now publish figures on how much of their own work it does. Those figures moved a lot this year. Most pages that explain the term were written before any of them existed.
Where did the idea of an intelligence explosion come from?
The term comes from the British mathematician I. J. Good. In a 1965 paper, Speculations Concerning the First Ultraintelligent Machine, he argued that a machine smarter than any person could design even better machines, and that there would then "unquestionably be an 'intelligence explosion'", leaving humans far behind.
Good's next line is the one people quote most. He called the first ultraintelligent machine "the last invention that man need ever make", and then added a condition: "provided that the machine is docile enough to tell us how to keep it under control." That one clause is pretty much the whole AI safety field, written sixty years early. You can read the quotes and their citation on Wikipedia's technological singularity page.
The science fiction writer Vernor Vinge carried the idea forward in 1993. He wrote that "within thirty years, we will have the technological means to create superhuman intelligence", and said he would be surprised if it came before 2005 or after 2030. We are now inside his window.
Most explainers stop about here. The Machine Intelligence Research Institute's Intelligence Explosion FAQ, still one of the top results for the term, describes itself as up to date as of 2013. It uses Deep Blue and IBM Watson as examples of the state of the art. It's good thinking. As evidence, it's badly dated.
How would an intelligence explosion actually work?
It works through returns on research. If each AI generation speeds up the work of building its successor by more than the next step gets harder, progress compounds and the gaps between releases shrink. If the steps get harder faster than the AI gets better, things slow down instead. The whole debate comes down to that ratio.
Here is a worked example. The numbers are made up to show the mechanism, not to forecast anything.
Say a lab needs 12 months of research to build model 2 with human researchers alone. Model 2 makes the research 1.5 times faster, but building model 3 is 1.2 times harder. The net speed-up is 1.5 / 1.2 = 1.25, so model 3 takes about 9.6 months. Keep going and the build times run 12, 9.6, 7.7, 6.1, 4.9 months. They shrink because the ratio sits above 1. That's an explosion in slow motion.
Now flip one number. Model 2 still makes research 1.5 times faster, but each step is 1.8 times harder. The ratio is 0.83, and generation times grow: 12, 14.4, 17.3 months. Same AI, same self-improvement, no explosion. Just a climb that keeps getting slower.
- Ratio above 1: each generation arrives sooner than the last. This is the "hard takeoff" picture.
- Ratio near 1: steady progress at a fixed pace, with AI doing more of the work.
- Ratio below 1: self-improvement happens, but it fizzles into diminishing returns.
Nobody knows the real ratio, which is why serious people disagree so sharply on timing. Labs have only just started publishing the data you'd need to estimate it.
Is recursive self-improvement already happening in 2026?
Partly. As of October 2026, AI writes most of the code and runs many of the experiments at the leading labs, but humans still set the research direction, judge results and decide what ships. Every major lab says it is not yet at full recursive self-improvement, meaning an AI that designs and builds its own successor without people.
This year's evidence is concrete.
Anthropic. In its report When AI builds itself, Anthropic says more than 80% of the code merged into its codebase was written by Claude as of May 2026. Before Claude Code's research preview in February 2025, that share was in the low single digits. The typical engineer merged 8 times as much code per day in Q2 2026 as in 2024. Anthropic adds that lines of code overstate the real gain. The report's own words: "recursive self-improvement is not inevitable", but it "could come sooner than most institutions are prepared for." AP, via Fortune, reported on 19 September 2026 that Claude now leads 26% of Anthropic's model research and development, under human supervision.
OpenAI. Last autumn, OpenAI set an internal goal of an automated AI research intern by September 2026. It says it hit that goal this September, according to Help Net Security's write-up of the announcement. The intern handles well-defined research tasks that would take a skilled researcher several days. But more than half of the successful 4 to 8 hour tasks still needed at least one human intervention. The next target is a fully automated AI researcher by March 2028. OpenAI's chief scientist Jakub Pachocki, in an essay titled An Alien Mind (6 September 2026), wrote that progress could continue into recursive self-improvement and called for "extreme caution", including voluntary slowdowns and safety bars checked by outside auditors.
Google DeepMind and xAI. Demis Hassabis calls the current stage "soft self-improvement", where AI helps people work smarter rather than becoming superintelligent itself, as Forbes reported on 7 October 2026. Elon Musk said in March that humans are "gradually getting less and less in the loop" on Grok, according to the AP report.
So the loop exists. People are still a required part of it, and every lab says so.
What are the signs an intelligence explosion is starting?
Watch for the time between capability jumps getting shorter, and for humans dropping out of the research loop. The clearest public signals in 2026 are task-length measurements from METR, the share of lab code and research done by AI, and how often AI agents still need a human to step in.
| Signal | Where it stands (latest public figure) | What would mean "speeding up" |
|---|---|---|
| Length of tasks AI can finish with 50% success | Doubling every ~7 months over 2019 to 2024, then ~4.3 months in METR's January 2026 update | The doubling time keeps shrinking |
| Longest task horizon measured | At least 16 hours for Claude Mythos (May 2026), the ceiling of METR's current test set | New tests needed just to measure it |
| Share of lab code written by AI | Over 80% of merged code at Anthropic (May 2026) | Code is no longer the bottleneck. Research judgment is |
| Share of model R&D led by AI | 26% at Anthropic (Sept 2026, AP) | Most research directions picked by the model |
| Human interventions needed | Over half of successful 4 to 8 hour tasks at OpenAI | Interventions become rare on multi-day work |
| Lab target for full automation | OpenAI: automated AI researcher by March 2028 | Targets pulled forward |
The METR figures come from its original study, Measuring AI Ability to Complete Long Tasks, which found the 7-month doubling, and the later updates summarized on METR's Wikipedia page. METR's own projection: if the trend holds, AI agents could carry out month-long projects on their own by the end of the decade.
Notice what the table shows. Code is mostly automated already. What's left is taste: picking the next problem, and knowing which results to trust. Anthropic's report flags human code review and research judgment as today's choke points.
Why might an intelligence explosion never happen?
Because the ratio in the worked example might stay below 1. Skeptics argue that each gain in intelligence takes more than a proportional amount of effort, that compute, chips and energy put hard limits on speed, and that AI still lacks the research creativity a runaway loop would need.
The strongest objections, each with a named source:
- Diminishing returns. Stuart Russell and Peter Norvig point out that most technologies follow an S-curve that speeds up and then levels off, not a curve that goes vertical. Ramez Naam argues higher intelligence may need much more than proportional effort.
- Physical limits. Anthropic's own report names chip fabrication, grid expansion and interconnect bandwidth as limits that could slow progress whatever the software does.
- Missing creativity. Anthropic co-founder Jack Clark has written about an "absence of valuable, intuitive creativity in today's AI systems", per Forbes. Gary Marcus argues current progress lets AI code faster, not get smarter.
- Slow measurement. A RAND report published 12 October 2026 notes that labs do not report automation shares in any consistent way. We can't tell how much of model improvement comes from AI-generated ideas, so nobody outside the labs can measure the ratio yet.
None of this rules it out. It just leaves the timing wide open. Anthropic's report puts three scenarios side by side: progress stalls, AI automates most development while humans steer, or full recursive self-improvement. It says the evidence points to the middle one for now.
Intelligence explosion vs singularity vs superintelligence: what's the difference?
An intelligence explosion is the process: a self-improvement loop that speeds up. Superintelligence is the result: an AI that beats the best humans at nearly every mental task. The singularity is the claimed aftermath: a point past which the future becomes impossible to predict. One can lead to the next, but each can be argued without the others.
| Term | What it describes | Who coined or popularized it |
|---|---|---|
| Intelligence explosion | The runaway self-improvement process | I. J. Good, 1965 |
| Recursive self-improvement (RSI) | The mechanism: AI improving AI | Lab and safety research term, now used by OpenAI and Anthropic |
| Superintelligence (SI, ASI) | The end state: intellect far beyond humans | Long-standing research term, now in lab names like Safe Superintelligence |
| Technological singularity | The unpredictable aftermath | Vernor Vinge, 1993 |
If you want the full breakdown of the labels, our guide What is SI (super intelligence)? covers how "SI" is used today, and SI vs AGI vs ASI separates the three acronyms people mix up.
What does an intelligence explosion mean for founders building AI products?
It means the model under your product will keep getting better on a shorter cycle, so anything you built to cover a model weakness has a shelf life. Build on what stays scarce as models improve: your users, your data, your distribution. And how often AI assistants recommend you, because they are becoming the place people ask first.
Two things change for you.
First, the gap between model releases becomes your planning unit. If task length doubles every four months or so, a workaround you ship in January for "the model can't do multi-step tasks" may be dead weight by summer. Say you run a small AI agent for bookkeeping and spent six weeks building a planner that splits jobs into small steps because the model kept losing the thread. Check every new model release against that planner. The day the raw model beats it, delete it.
Second, people now take questions about this topic, and about your product category, to AI assistants first. When someone asks ChatGPT or Perplexity what an intelligence explosion is, the answer cites a handful of pages. The same thing happens when they ask which tool to use for your problem. Being one of those cited pages is a distribution channel, and most analytics tools can't even see it.
Do this tonight
- Read Anthropic's When AI builds itself report in full. It's the most detailed primary data published so far on AI doing AI research.
- Write down the one feature in your product that exists only because today's model is weak at something. Put a calendar reminder to retest it on the next major model release.
- Ask ChatGPT, Claude and Perplexity the three questions your product is meant to answer. Note whether you appear, and which pages get cited instead.
- Bookmark METR's time-horizon page and check the doubling time each quarter. If it keeps shrinking, shorten your roadmap horizon with it.
If you want to see whether those assistants are already sending you visitors, Clerion reports ChatGPT, Perplexity, Claude, Gemini and Copilot as their own channel, so that traffic stops hiding in "direct".
Related reading
- What is SI (super intelligence)?: what the term means and why founders use it.
- What is superintelligence, and how close are we?: the forecasts and who makes them.
- Is superintelligence dangerous? What the evidence says: the risk case, weighed.
Frequently asked questions
Who coined the term intelligence explosion?
The British mathematician I. J. Good, in his 1965 paper Speculations Concerning the First Ultraintelligent Machine. He argued a machine smarter than any person could design better machines, setting off an intelligence explosion.
Is an intelligence explosion the same as the singularity?
No. An intelligence explosion is the process of AI improving itself faster and faster. The technological singularity, popularized by Vernor Vinge in 1993, is the claimed aftermath: a point past which the future can't be predicted.
What is recursive self-improvement?
Recursive self-improvement (RSI) is AI helping to build better versions of itself. Anthropic describes the full version as an AI that can fully autonomously design and develop its own successor. No lab says it has reached that as of October 2026.
How much of AI lab code is written by AI in 2026?
Anthropic says Claude wrote more than 80% of the code merged into its codebase as of May 2026, up from low single digits before February 2025. AP reported in September 2026 that Claude leads 26% of Anthropic's model R&D.
When could an intelligence explosion happen?
Nobody knows. OpenAI targets a fully automated AI researcher by March 2028, and METR's task-length doubling time fell to about 4.3 months in early 2026. Skeptics point to diminishing returns and limits on chips and energy.
What signs would show an intelligence explosion is starting?
Shrinking gaps between model generations, a falling doubling time on METR's task-length measure, AI picking most research directions rather than only writing code, and AI agents needing human help less often on multi-day tasks.
Sources
- Technological singularity (I. J. Good 1965, Vinge 1993 quotes)
- MIRI: Intelligence Explosion FAQ
- Anthropic: When AI builds itself
- AP via Fortune: AI companies get closer to recursive self-improvement (19 Sept 2026)
- Help Net Security: OpenAI just hit a milestone on the road to self-improving AI (7 Sept 2026)
- OpenAI: An Alien Mind, by Jakub Pachocki
- Forbes: What happens when AI learns to improve itself? (7 Oct 2026)
- METR: Measuring AI Ability to Complete Long Tasks
- METR (Wikipedia, time horizon updates)
- RAND: Recursive Self-Improvement and Automated AI Research and Development (12 Oct 2026)