Superintelligence is an intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest. That is the definition philosopher Nick Bostrom gave in his 2014 book Superintelligence, and it remains the standard one. Note what it does not say: it does not say “a robot,” “a god,” or “a computer that thinks like a human, only faster.” It describes a level of capability — better than the best humans at essentially every kind of thinking — without saying what form it takes.
The short version: if AGI is an AI as smart as a human across the board, superintelligence is an AI vastly smarter than the smartest humans at everything, including at improving itself. That last clause is what makes the idea so consequential, and so contested.
How it differs from today’s AI
Today’s AI is narrow: brilliant within a boundary, helpless outside it. A chess program beats every human on Earth and cannot make a sandwich. A language model can draft a legal brief and cannot tell whether its own argument is true. Each system is built and trained for a class of tasks, and its competence ends at the edge of that class.
Superintelligence is the opposite of narrow. It would be general — able to take on new problems the way a good human researcher does — and then better than any human at them: faster, with a vastly larger working memory, able to run thousands of copies of itself in parallel, and able to learn from every copy at once. The philosopher David Chalmers breaks the path into three steps: first AI reaches human equivalence, then it is extended beyond human ability, then amplified until it dominates humans across arbitrary tasks.
A helpful image comes from Anthropic’s Dario Amodei: a “country of geniuses in a datacenter.” He imagines millions of copies of a system smarter than a Nobel Prize winner across most fields — biology, programming, mathematics, engineering, writing — each able to work autonomously on tasks lasting days or weeks, running at roughly ten to a hundred times human speed, and collaborating the way human researchers do. That is not one superintelligent mind. It is a civilization-scale workforce that happens to be software.
How we might get there
The most discussed route is recursive self-improvement: an AI system good enough at AI research to improve its own design, producing a smarter successor, which improves itself further, and so on. Each step makes the next step easier. Bostrom called the result an intelligence explosion. The writer Vernor Vinge predicted in 1993 that within thirty years we would have “the technological means to create superhuman intelligence.”
There are other routes. A system could become superintelligent by sheer scale — more computing power, more data, better algorithms — without ever redesigning itself. Or many merely human-level systems could be combined into a collective superintelligence: a large number of reasoning systems that communicate and coordinate well enough to act, in aggregate, far beyond any individual. Bostrom also notes the simplest route of all: take a human-like mind and run it on much faster hardware. Biological neurons fire at about 200 times a second; a modern processor runs at around 2 billion cycles a second. A human-level thinker running a million times faster would be, for most practical purposes, superintelligent already.
None of these routes is proven. The disagreement is over which, if any, will actually happen — and how fast.
What would it be like to live with?
Amodei’s essay is the most concrete attempt to answer this, and it is worth reading because it resists both the utopian and the apocalyptic reflex. His picture of powerful AI includes five areas of transformation: biology and health, neuroscience and mental health, economic development, peace and governance, and work and meaning. The mechanism is not magic; it is the application of enormous intelligence to problems that are currently limited by how many smart people can work on them.
But Amodei also lists the factors that would limit even superintelligence. Intelligence is not the only ingredient in getting things done. The outside world moves at its own speed: cells grow, experiments take weeks, hardware must be built. Data is sometimes missing, and no amount of cleverness substitutes for it. Some systems are intrinsically unpredictable. Human institutions — laws, clinical trials, politics — constrain what can be done even by a willing genius. And physical law sets hard ceilings. His conclusion is that the truth is “some messy admixture” of instant transformation and slow grind, varying by field.
That matters because the popular image of superintelligence — a single mind that solves everything overnight — is the least likely version. A superintelligence would be constrained by the world it operates in, even as it reshapes that world far faster than we could.
Why people argue about it
Three questions divide the field:
Is it inevitable? Some argue that once human-level AI exists, the step to superintelligence follows from the logic of improvement itself. Others argue that intelligence has natural ceilings, or that the hard problems are not intelligence problems at all — they are problems of data, energy, and the physical world.
Is it one thing or many? The public debate often imagines a single superintelligence, but the realistic versions are plural: many systems, many copies, many owners. That changes the policy problem completely. Governing one system is different from governing an ecosystem.
Is it near or far? In 2023, OpenAI’s leaders Sam Altman, Greg Brockman, and Ilya Sutskever wrote that superintelligence “may happen in less than 10 years.” In October 2025, hundreds of public figures signed an open letter calling for a ban on its development. The disagreement is not only about dates; it is about whether to treat superintelligence as a research problem, a governance problem, or an emergency.
What the scoreboard says
SiPoly tracks superintelligence claims under its AGI & superintelligence timelines topic, and the record illustrates the definitional problem. Sam Altman’s September 2024 claim that “it is possible that we will have superintelligence in a few thousand days” was listed as too vague to grade: with no definition and no date, there was nothing to test (claim page). By contrast, Amodei’s “country of geniuses in a data center” is gradeable, precisely because he defined it in detail — smarter than top human experts across most fields, working autonomously on tasks lasting days or weeks, in many parallel copies — and gave it a deadline of September 2028 (claim page).
Elon Musk’s June 2025 claim of “digital superintelligence” in 2025, “or next year for sure,” is open and due at the end of 2026. The lesson the scoreboard keeps teaching is that superintelligence is easy to invoke and hard to specify — and that the claims worth tracking are the ones that say what would count.
Across the whole scoreboard, as of 5 November 2026, 367 claims are tracked, 129 resolved, and 60% came true.
Key sources
- Bostrom, N., Superintelligence: Paths, Dangers, Strategies (Oxford University Press, 2014) — the standard definition and the fullest treatment of the paths to superintelligence. Summary and context: https://en.wikipedia.org/wiki/Superintelligence
- Amodei, D., “Machines of Loving Grace” (October 2024) — the “country of geniuses in a datacenter” and the limits on what intelligence alone can do. https://darioamodei.com/machines-of-loving-grace
- Altman, S., “The Intelligence Age” (September 2024) — “It is possible that we will have superintelligence in a few thousand days.” https://ia.samaltman.com/
- Kokotajlo, D. et al., AI 2027 — a scenario that ends in artificial superintelligence in December 2027. https://ai-2027.com/
Common misconceptions
- “Superintelligence means a single godlike robot.” It could equally be a distributed system, a collective of many copies, or a human-like mind running far faster than a brain. The definition is about capability, not form.
- “Superintelligence is just a much smarter chatbot.” A chatbot answers questions. A superintelligence would set its own sub-goals, carry out multi-week projects, design experiments, and improve its own successors. The difference is autonomy and generality, not just IQ points.
- “It’s science fiction, decades away at least.” That was the consensus a decade ago. Today the leaders of the major labs put it within years, and the question is argued in the UN Security Council and in national legislatures. You can disagree with the timeline without dismissing the topic.