The Last Invention/Superintelligence
What is superintelligence?
Superintelligence is a hypothetical form of artificial intelligence that outperforms the best human minds at practically everything. Here is what the term means, where it came from, and why serious people argue about it.
Key points
- Superintelligence means AI that exceeds human ability across virtually all intellectual work, not just one task.
- It is a step beyond artificial general intelligence (AGI), which means roughly human-level ability.
- It does not exist today. Estimates of when, or whether, it will arrive vary enormously.
- The idea dates to the 1950s and 1960s. The word itself was popularised by Nick Bostrom’s 2014 book.
The short definition
The most cited definition comes from the philosopher Nick Bostrom, who described a superintelligence as:
“any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest.”
Two words in that sentence do the work. Greatly rules out a system that is merely a little better than a clever person. Virtually all rules out specialists. A chess engine has beaten every human for decades, but it cannot plan a holiday or write a paragraph. A superintelligence would be better than the best people at science, engineering, strategy, writing and persuasion, all at once.
You will also see it written as ASI (artificial superintelligence) or simply SI.
AI, AGI and superintelligence
The three terms describe a ladder of capability.
| Term | What it means | Status |
|---|---|---|
| Narrow AI | Performs one kind of task, sometimes far better than people. Chess engines, translation systems and protein-structure predictors are examples. | Everywhere |
| Artificial general intelligence (AGI) | Matches a capable person across most intellectual tasks. | Disputed; not yet by most definitions |
| Superintelligence (ASI, SI) | Exceeds the best people across virtually all of them. | Hypothetical |
None of these lines is sharp. There is no agreed test for AGI, and today’s large language models blur the picture: they are far more general than a chess engine, yet uneven in ways no human expert is.
Three ways to be superintelligent
Bostrom separates three forms, which can overlap.
- Speed. A mind that thinks as we do, only much faster: a researcher who gets a thousand years of thinking done in one.
- Collective. A large number of human-level minds that coordinate well enough to outperform any organisation of people.
- Quality. A mind that is not just faster but better, able to grasp things we cannot, roughly as we grasp things other animals cannot.
Where the idea came from
- 1951. Alan Turing says in a talk that once machines can think, it will probably not take them long to outstrip us.
- 1965. I. J. Good defines the “ultraintelligent machine” and argues it would set off an intelligence explosion.
- 1993. The mathematician and novelist Vernor Vinge predicts superhuman intelligence within thirty years and names the moment the technological singularity.
- 2014. Bostrom’s book makes superintelligence the standard term and turns the question of controlling it into a research agenda.
The longer story is in the AI timeline.
Does superintelligence exist?
No. The most capable AI systems today can write, program, summarise research and solve many problems that once needed an expert. They also make mistakes no expert would make, and they depend on people to set goals, check results and build the next version.
Opinions on timing run from “within a few years” to “not this century” to “never in the form people imagine”. Surveys of AI researchers show wide disagreement, and forecasts have moved a long way in both directions. Treat any confident date with suspicion.
Why it matters
If it can be built, superintelligence would be the most consequential technology in history, for the reason Good gave: it would be able to do the inventing. Cures, clean energy and scientific questions that have resisted us for generations might yield quickly.
The same capability is the risk. A system much more capable than its makers is hard to correct if what it pursues differs, even slightly, from what they intended. Making sure advanced AI does what people actually want is known as the alignment problem, and it is unsolved.
Keep reading
The intelligence explosion
The 1965 argument that a smart enough machine would improve itself, the case for it, and the reasons to doubt it.
Read the argument →A timeline of AI
From a 1943 model of the neuron to chatbots with hundreds of millions of users: the dates that mattered.
See the timeline →