What is an intelligence explosion?
How the loop works
- Automate AI research. An AI system becomes useful at the work AI researchers do: running experiments, writing training code, finding better algorithms.
- Run it at scale. Unlike human researchers, it can be copied thousands or millions of times and run faster than real time.
- Feed the gains back. Better algorithms and methods produce a more capable model, which is a better researcher.
- Repeat. If each cycle gives more improvement than the last, progress speeds up rather than levelling off.
Leopold Aschenbrenner's 2024 essay Situational Awareness put this in concrete terms: hundreds of millions of AGI-level researchers could fit a decade of algorithmic progress into less than a year. The AI 2027 scenario (2025) imagined a similar path ending in superintelligence before 2030.
Fast vs slow takeoff
| Fast takeoff | Slow takeoff | |
|---|---|---|
| AGI to ASI | Days to months | Years to decades |
| Main driver | Software self-improvement | Compute, energy and experiments |
| Warning time | Little | Enough to adapt and regulate |
| Implication | Safety has to be solved in advance | Mistakes can be fixed along the way |
What could slow it down
- Compute. Better algorithms still need chips and power to train on, and both take years to build.
- Diminishing returns. The easy improvements get found first, and each later gain may be harder.
- Verification. Checking that a better-than-human result is actually correct is hard. DeepMind's 2026 From AGI to ASI report flags this as a key friction.
- Real-world experiments. Biology, materials and robotics need physical testing that can't be run at machine speed.
The 2026 evidence
For years this was theory. In 2026 it became a question about real systems:
- DeepMind's June 2026 report listed recursive self-improvement as one of four pathways from AGI to ASI.
- OpenAI reported that an internal long-horizon model found an improvement to its training process on its own.
- On July 27, 2026, Sam Altman said AI had crossed the threshold of being able to improve itself. This is a company claim without independent verification.
- OpenAI is publicly aiming for an automated AI researcher by 2028.
Each development is logged with sources on the tracker. For what the loop would mean for safety, see Is superintelligence dangerous?
Frequently asked questions
Who came up with the intelligence explosion?
I. J. Good, a British statistician who had worked with Alan Turing. In 1965 he argued that an 'ultraintelligent machine' could design better machines, setting off an intelligence explosion, and that such a machine would be the last invention humanity needed to make.
What is recursive self-improvement?
The mechanism behind the explosion: an AI system improving the processes that produce AI, such as algorithms, training methods, code or chip design, so that each generation is better at improving the next.
What is the difference between fast and slow takeoff?
In a fast takeoff, the step from human-level to vastly superhuman takes days to months, leaving little time to react. In a slow takeoff it takes years, giving society time to adapt and correct course. Researchers disagree sharply about which is more likely.
Has the intelligence explosion started?
That's disputed. In July 2026 OpenAI's CEO claimed AI had crossed the threshold of being able to improve itself, and OpenAI reported a model finding a training improvement on its own. Independent evidence of a compounding, runaway loop has not been shown.