5 October 2026 · 2 min read
The intelligence explosion, six decades on
In 1965 a statistician described a machine that improves itself. The argument has barely changed since. The evidence has.
I. J. Good worked with Alan Turing at Bletchley Park during the war. In 1965 he published a paper on what he called the ultraintelligent machine: one that could outdo any person at every intellectual activity. Designing machines is an intellectual activity, he reasoned, so such a machine could design a better one, which could design a better one again. He named the result an intelligence explosion and suggested it would be the last invention people would need to make themselves.
Vernor Vinge, a mathematician and novelist, gave the idea its popular name in a 1993 essay on the technological singularity. Both texts share the same structure: once the thing doing the research is also the thing being improved, progress feeds itself.
The loop, stated plainly
- AI systems become useful for AI research: writing code, running experiments, proposing ideas.
- Research goes faster, so better systems arrive sooner.
- Better systems are more useful for research. Return to step one.
This is called recursive self-improvement. Nothing in the loop requires a mysterious leap. It requires only that each turn of the cycle yields enough improvement to keep the next turn going.
What could slow it down
The strongest objections are practical. Training a frontier model needs chips, power and buildings, and those are produced on industrial timescales, not software ones. Many experiments cannot be rushed: a training run takes as long as it takes. Easy improvements get found first, so each further gain may cost more than the last. And a system that is excellent at code is not automatically excellent at deciding which research question is worth asking.
What could speed it up
Software has no factory. An algorithmic improvement can be copied to every machine the same day. A capable research system can run as thousands of instances in parallel, which is something no human laboratory can do with its best scientist. If ideas are the bottleneck and ideas can be automated, the brakes above matter less than they appear to.
What to watch
The debate between a slow and a fast takeoff will not be settled by argument. It will be settled by one measurable quantity: how much of the work of AI research is done by AI systems, and how that share changes from year to year. A share that creeps up means the loop is weak. A share that climbs quickly means Good was describing the present.
Sources: I. J. Good, "Speculations Concerning the First Ultraintelligent Machine" (1965); Vernor Vinge, "The Coming Technological Singularity" (1993).