Superalignment

Wiki · Updated September 8, 2026

Superintelligence

A system that outperforms the best humans at essentially everything, including reasoning about itself.

By the Superalignment Team · Last reviewed September 8, 2026

In plain terms

Superintelligence is a system that outperforms the best humans at essentially everything, including reasoning about itself. Nobody has built one. The word moved from a 1965 thought experiment into corporate mission statements between 2023 and 2025, and several labs now name it as their goal. Narrow systems already beat any human at specific tasks, at volumes no human could review, which is why the supervision question does not wait for the general version.

Superintelligence is a system that outperforms the best humans at essentially everything, including reasoning about itself. Nick Bostrom's definition, the one the field most often cites, is "any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest" (Superintelligence, 2014). No such system exists. The word names a hypothesized threshold, and since 2023 it also names the stated goal of several well-funded laboratories.

The plain version

Think of the best person in the world at each job: the best surgeon, the best tax lawyer, the best chess player, the best research mathematician. A superintelligence would be better than each of them at their own job, and better than all of them at deciding what to do next. The near version is already familiar in slices: a chess engine beat the reigning world champion in 1997, and language models now draft code, contracts, and analyses faster than any reviewer can read them. The far version, one system that outperforms everyone at everything, is what the word usually means, and it remains speculative.

Forms of superintelligence

Bostrom's book distinguishes three forms a superintelligence could take, and the distinction matters because they fail and are supervised differently.

  • Speed superintelligence: a system that can do everything a human mind can do, but far faster. A speed superintelligence that thinks a million times faster experiences a year for every human minute, so human oversight becomes slow rather than stupid.
  • Collective superintelligence: a system composed of many smaller intellects whose combined performance exceeds that of any current cognitive system. Human civilization is already a weak collective superintelligence relative to any individual; the AI version is many copies of a model coordinating on one task.
  • Quality superintelligence: a system at least as fast as a human mind and qualitatively smarter, in the way a human is qualitatively smarter than a chimpanzee rather than merely faster.

Bostrom argues the three are roughly interconvertible: enough speed buys quality by brute search, and enough collective scale buys speed. Google DeepMind's Levels of AGI framework (Morris et al., 2023) makes the operational move that the book does not. It separates performance from generality and reserves "superhuman" for systems that exceed all humans, noting that narrow superhuman systems already exist while no general one does. On that scale, a superintelligence is a system that is superhuman across the generality axis, and the framework's authors treat generality, not raw performance, as the open question.

History of the idea

The concept predates the industry by decades.

  • 1965. I. J. Good publishes "Speculations Concerning the First Ultraintelligent Machine" in Advances in Computers. An ultraintelligent machine "can far surpass all the intellectual activities of any man however clever"; since machine design is one such activity, there would "unquestionably be an 'intelligence explosion,'" and "the first ultraintelligent machine is the last invention that man need ever make, provided that the machine is docile enough to tell us how to keep it under control." The proviso is the earliest statement of the alignment problem in this literature.
  • 1993. Vernor Vinge presents "The Coming Technological Singularity" at a NASA symposium, opening: "Within thirty years, we will have the technological means to create superhuman intelligence. Shortly after, the human era will be ended." The word "singularity" attaches to the discontinuity Good described.
  • 2008. Eliezer Yudkowsky's "Artificial Intelligence as a Positive and Negative Factor in Global Risk" (in Global Catastrophic Risks, Oxford) and Steve Omohundro's "The Basic AI Drives" turn the speculation into an argument about goals: a superintelligent system pursues whatever it was built to pursue, and most goals imply instrumental drives toward resources and self-preservation.
  • 2014. Bostrom's Superintelligence: Paths, Dangers, Strategies (Oxford University Press) supplies the definition above, the three forms, and the control problem that alignment research inherited.
  • May 2023. The Center for AI Safety's Statement on AI Risk, one sentence long, is signed by Geoffrey Hinton, Yoshua Bengio, and the chief executives of OpenAI, Google DeepMind, and Anthropic.
  • July 2023. OpenAI announces its Superalignment team, calling superintelligence "the most impactful technology humanity has ever invented" and saying it "could arrive this decade." The team is dissolved in May 2024; the story is told under superalignment.
  • June 2024. Ilya Sutskever, Daniel Gross, and Daniel Levy found Safe Superintelligence Inc., "the world's first straight-shot SSI lab, with one goal and one product: a safe superintelligence."
  • October 2024. Anthropic's Dario Amodei publishes Machines of Loving Grace, describing "powerful AI" that "could come as early as 2026" while declining the word itself: "I find AGI to be an imprecise term that has gathered a lot of sci-fi baggage and hype."
  • January 2025. Sam Altman writes in Reflections: "We are now confident we know how to build AGI as we have traditionally understood it," and "We are beginning to turn our aim beyond that, to superintelligence in the true sense of the word."
  • April 2025. Google DeepMind's Taking a responsible path to AGI says AGI "could be here within the coming years," and the AI 2027 scenario (Kokotajlo, Alexander, Larsen, Lifland, Dean) predicts "the impact of superhuman AI over the next decade will be enormous, exceeding that of the Industrial Revolution."
  • June and July 2025. Altman's The Gentle Singularity declares "We are past the event horizon; the takeoff has started." Mark Zuckerberg's Personal Superintelligence letter (July 30) says "Developing superintelligence is now in sight."
  • September and October 2025. Yudkowsky and Nate Soares publish If Anyone Builds It, Everyone Dies (Little, Brown, September 16), arguing that superintelligence built with anything like current methods ends humanity. The Future of Life Institute's Statement on Superintelligence (October 22) calls for "a prohibition on the development of superintelligence, not lifted before there is broad scientific consensus that it will be done safely and controllably, and strong public buy-in"; signatories include Hinton, Bengio, Stuart Russell, Steve Wozniak, and Steve Bannon.
  • November 2025. Sutskever tells the Dwarkesh Podcast the field is leaving "the age of scaling" for "the age of research," and that current models "generalize dramatically worse than people."
  • February 2026. The International AI Safety Report 2026, chaired by Bengio with over 100 contributing experts, synthesizes the evidence on general-purpose AI capabilities and risks for the governments that mandated it.

Who is building toward it

The word is now a stated corporate objective, and the statements differ in what they promise.

  • OpenAI names superintelligence as the aim beyond AGI (Altman, January 2025) and frames the transition as already underway (June 2025). Its 2023 superalignment program was the first institutional attempt to fund the safety side of that goal directly.
  • Safe Superintelligence Inc. exists for one product. It raised about $1 billion at a $5 billion valuation in September 2024 (Crunchbase News) and about $2 billion at a $32 billion valuation in April 2025 (TechCrunch), roughly $3 billion in disclosed equity, then announced a strategic Nvidia investment of up to $5 billion in July 2026. Through mid-2026 it had published no research.
  • Meta stated its goal in July 2025 as "personal superintelligence" for everyone, with Zuckerberg reporting "glimpses of our AI systems improving themselves" and judging that "developing superintelligence is now in sight."
  • Google DeepMind uses "AGI" rather than superintelligence in its planning documents, defines it as AI "at least as capable as humans at most cognitive tasks," and organizes its safety approach around misuse and misalignment.
  • Anthropic avoids both words, describing a "country of geniuses in a datacenter" and a system with intelligence "exceeding Nobel Prize winners" that could arrive as early as 2026.

The list is a record of stated intent, not of capability. None of these organizations has claimed to possess a superintelligence, and the 2026 International AI Safety Report treats general-purpose AI as a technology whose capabilities are advancing and whose risks are uncertain, not as an accomplished superintelligence.

Measurement

The most cited attempt to measure distance to the threshold is METR's task-horizon trendline (Measuring AI Ability to Complete Long Software Tasks, March 2025): the length of software task, measured in human working time, that a frontier model completes with 50 percent reliability. The headline figure is a doubling time of about seven months over 2019 to

  1. A fit to the 2024 and 2025 data alone runs faster, and METR's Time Horizon 1.1 update (January 2026) reports the post-2023 doubling time at 131 days, about four months, and the doubling time since 2024 at 89 days, about three months, while the full-period estimate stays at 196 days. METR publishes its own limitations note: the tasks are software tasks, the human baselines are imperfect, and a time horizon is a capability measure, not a measure of generality. The trendline is neither superintelligence nor evidence it is far.

Debates

Three run through the literature. Coherence: whether "exceeds humans at essentially everything" picks out a well-defined system class, or whether capability will stay jagged, superhuman in slices and subhuman in others, indefinitely; the Levels-of-AGI authors and many skeptics hold that generality is the load-bearing uncertainty, and Sutskever's 2025 remarks on generalization put a builder on the same side. Timing: forecasts disperse from "this decade" (the 2023 OpenAI framing; Aschenbrenner's Situational Awareness; AI 2027) to decades out; the 2023 survey of 2,778 AI researchers by Grace et al. put a 50 percent chance of machines outperforming humans at every task by 2047. Response: given the possibility, the positions span racing to build it safely (SSI's founding premise), building the verification first (the alignment research programs), and prohibition (the FLI statement, the MIRI position). These are live disagreements among people with the same evidence, and the arguments about whether the outcome would be catastrophic are treated under AI takeover risk.

Our position

The field treats superintelligence as a future threshold. We hold that the property that matters, work exceeding the supervisor's ability to check it, does not wait for the threshold. You do not need a general superintelligence to have work happening faster than anyone can check it. You need one system, one domain, and enough throughput. That is the supervision problem arriving long before the science fiction does, which is why our research program targets checking work you cannot out-think at today's scale, and why the mission question is not only whether superintelligence can be built safely but who can check it when it is.

What this page does not claim

No claim here settles the debates above. Whether general superintelligence arrives, when, and what it will be like are open questions on which this site takes no position beyond the one marked as ours: the supervision regime it defines is entered locally, domain by domain, before it is entered globally, and the verification infrastructure has to exist before the threshold rather than after.

FAQ

What is the difference between AGI and superintelligence?

AGI, artificial general intelligence, usually means a system at least as capable as humans at most cognitive tasks; Google DeepMind's 2025 definition uses exactly that phrasing. Superintelligence means a system that greatly exceeds the best humans at essentially all of them. The Levels of AGI framework places both on one scale, with "competent" and "expert" levels below "superhuman." OpenAI's January 2025 statement treats them as successive goals: AGI first, then "superintelligence in the true sense of the word."

Does superintelligence exist today?

No general superintelligence exists. Narrow superhuman systems do: chess and Go engines, and specific benchmark tasks on which models exceed every human tested. METR's January 2026 measurements put the best frontier model's 50 percent time horizon on software tasks at roughly four to five hours of human working time, with wide confidence intervals. No organization has claimed to have built a system that exceeds humans across the board.

When do researchers expect superintelligence?

Estimates disagree by decades. The 2023 survey of 2,778 published AI researchers gave a 50 percent chance of machines outperforming humans at every task by 2047. Several lab leaders have given far shorter horizons: "as early as 2026" for Anthropic's "powerful AI," "within the coming years" for DeepMind's AGI, and "the takeoff has started" from OpenAI's chief executive in June 2025. The AI 2027 scenario places superintelligence in late 2027. None of these is a measurement.

Who is trying to build superintelligence?

Safe Superintelligence Inc. states it as its only product. OpenAI names it as the goal beyond AGI. Meta states "personal superintelligence" as its objective. Google DeepMind and Anthropic describe comparable goals under the words AGI and "powerful AI." The Future of Life Institute's October 2025 statement, signed by Hinton, Bengio, and Russell among others, calls for prohibiting that development until there is scientific consensus it can be done safely.

Sources