Superintelligence: what the word means, and why it is everywhere now

Two years ago, superintelligence was mostly a word from philosophy books. Today it is in the names of research labs, in the mission statements of the largest technology companies, and, since October 4, in the name of a US federal task force. A word used this often starts to lose its shape. This article sets out what it means, where the idea came from, and who is now building toward it.

A working definition

The most cited definition comes from the philosopher Nick Bostrom, in his 2014 book Superintelligence. He describes it as "any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest."

Two parts of that sentence do the work. "Greatly" means far beyond the best people, not slightly better. "Virtually all domains" means across the board. A chess engine is far better than any human at chess, but it is not superintelligent, because chess is one domain. Artificial general intelligence, or AGI, usually means a system that matches people across most domains. Superintelligence is the step beyond that.

The oldest argument

The idea is older than modern AI. In 1965 the statistician I. J. Good, who had worked with Alan Turing at Bletchley Park, published an essay on what he called an ultraintelligent machine. His argument was short. Designing machines is an intellectual activity. A machine better than us at every intellectual activity could therefore design a better machine than we can, which could design a better one again. Good called the result an "intelligence explosion" and concluded that the first such machine would be "the last invention that man need ever make."

That loop, a system improving the systems that come after it, is still the core of most serious thinking about superintelligence. It is also why researchers now watch one capability closely: how much of AI research itself AI can do.

Three ways to be smarter

Bostrom also separates kinds of superiority. A speed superintelligence thinks like a person, only far faster. A collective superintelligence is many systems working together, so that the group is far more capable than any one of them. A quality superintelligence thinks better, not just faster or in greater numbers. Real systems may mix all three. Today's AI already shows the first two in narrow ways: it answers in seconds, and thousands of copies can run at once.

Who is building toward it

The shift in language is easy to date.

• In July 2023, OpenAI announced a superalignment team and wrote that superintelligence "could arrive this decade."
• In June 2024, Ilya Sutskever, OpenAI's former chief scientist, co-founded Safe Superintelligence Inc., which calls itself the world's first "straight-shot SSI lab". In July 2026, Nvidia announced a strategic partnership with the company.
• In June 2025, Mark Zuckerberg told Meta staff that "developing superintelligence is coming into sight" and formed Meta Superintelligence Labs.
• The same month, Sam Altman wrote: "We are past the event horizon; the takeoff has started."
• In November 2025, Microsoft AI formed a superintelligence team under Mustafa Suleyman, aiming at what he calls humanist superintelligence: advanced AI kept under human control and built to solve concrete problems, starting with expert-level medical diagnosis.

And who wants to wait

Not everyone agrees the race should continue. In October 2025, a Statement on Superintelligence hosted by the Future of Life Institute called for a prohibition on developing superintelligence until there is broad scientific consensus that it can be done safely and controllably, and strong public buy-in. Its signers include the AI pioneers Geoffrey Hinton and Yoshua Bengio, the computer scientist Stuart Russell and Apple co-founder Steve Wozniak.

How we read it

No system today meets Bostrom's definition, and nobody can tell you the date one will. Forecasts vary widely, and the people building these systems disagree with each other. What is clear is the direction: more of the world's judgment work is moving to machines, and the systems doing it are getting more capable each year.

The harder question is not whether AI will get smarter. It is whether people will still be able to tell when it is right.

That question runs through this research track. The next articles look at how close we are, how systems that outgrow their teachers can stay on course, how people can check work they could not do themselves, what we can see inside models, and what governments are starting to do.