The short version
Superintelligence means an artificial system that performs far better than the most capable humans at virtually every intellectual task, including scientific research, planning and persuasion. No such system exists today. Whether one can be built, and when, is the most consequential open question in technology.
Three rungs on a ladder
- Narrow AI does one kind of task, sometimes far better than people. A chess engine and a protein structure predictor are narrow systems.
- Artificial general intelligence (AGI) would handle most intellectual work at the level of a capable person, across fields.
- Superintelligence would go well past the human level across the board.
The difference between the last two is covered in AGI and superintelligence are two different claims.
How it could happen
Three routes are discussed seriously. The first is scale: today's methods keep improving as they are given more computing power, more data and better algorithms. The second is self-improvement: AI systems start doing AI research and speed up their own progress. The third is a new idea that nobody has had yet. The first two are explained in The three inputs and The intelligence explosion, six decades on.
When
Nobody knows. Informed forecasts range from a few years to many decades, and some researchers doubt it will happen at all. The spread is the finding. A reader should distrust any source that gives a single confident date.
Why it matters
A system that outperforms people at research could speed up medicine, energy and science by amounts that are hard to picture. The same capability, pointed at the wrong goal or controlled by too few people, could cause harm on the same scale. The field that works on making such systems do what is intended is called alignment. Start with Alignment is an engineering problem first.
Where to go next
Read the journal in order, keep the glossary open in another tab, and see the timeline for how the idea developed. Our own position is set out in the manifesto.