The intelligence explosion: how AGI could turn into superintelligence
If AI gets good enough to improve AI, progress could speed up dramatically. How likely that is may matter more than when AGI arrives.
Published 6 October 2026 by Omar Nouha Sane
The idea
The concept goes back to I. J. Good in 1965. A machine that is better than humans at designing machines could build an even better successor, which could build a better one again. Each step would come faster than the last. Good called this an intelligence explosion.
Today the question is usually framed as “takeoff speed”: once AI reaches roughly human level, how long until it is far beyond human level?
Arguments for a fast takeoff
- AI doing AI research. Leading labs already use AI to help write code and run experiments. If systems become as good as top researchers, the number of “researchers” could jump from thousands to millions.
- Copies and speed. A trained model can be copied instantly and can run faster than a human thinks. Human institutions cannot grow like that.
- No biological limits. Human brains are limited by size and energy. Software can be scaled with more hardware.
Arguments for a slow takeoff
- Compute bottlenecks. Better ideas still need to be tested on large amounts of computing power, which takes chips, power plants and time to build.
- Diminishing returns. Each improvement might be harder to find than the last, cancelling out the speed-up.
- The real world is slow. Many advances need physical experiments, factories or regulation, which do not move at software speed.
- Gradual is already happening. Progress so far has come in steady steps across many labs, not as one sudden jump.
Why it matters
A slow takeoff would give society years to notice problems, adapt rules and fix mistakes. A fast one might leave weeks or months. That difference is why many people concerned about AI risk focus as much on takeoff speed as on the arrival date. It also connects to alignment: the less time there is, the more important it is that systems are aligned from the start.
What markets say
Two markets on the tracker deal with this directly. One asks whether there will be less than a year between the first AGI and the first superintelligence, the other whether there will be an intelligence explosion before 2035. You can see the current values in the market list. They are worth comparing: traders can expect a fast jump in principle while still doubting it happens soon.