The core guide
What superintelligence is, how it differs from AI and AGI, how and when it might arrive, the risks and attempts to control it, the organisations building frontier AI, and a glossary.
The basics
What is superintelligence?
The established meaning of the word in AI research, where it comes from, why it is not a description of today's AI, and why it matters more now that the word is in the news.
Last reviewed 22 Sept 2026
Is AI the same as superintelligence?
No. "Super Intelligence" is the US administration's new name for AI in general; "superintelligence" in AI research means something that does not exist yet. Here is the difference, in plain English.
Last reviewed 22 Sept 2026What is AGI?
Artificial general intelligence, in plain English: what the term means, how it differs from today's AI and from superintelligence, and why the US administration's "Super Intelligence" is neither.
Last reviewed 23 Sept 2026What is an AI agent?
What turns an AI model into an agent, how much today's agents can really do on their own, how they fail, why permissions and sandboxes matter, and why agents sit at the centre of debates about AGI and superintelligence.
Last reviewed 3 Oct 2026AGI vs superintelligence: what's the difference?
The two upper rungs of the AI ladder, told apart in plain English, and why September 2026 made the words harder to use.
Last reviewed 23 Sept 2026
How would we know if AGI had arrived?
Why there is no agreed test for artificial general intelligence: the Turing test, benchmarks and why they saturate, levels of AGI, autonomy and economic tests, and how to read a claim that AGI is here.
Last reviewed 3 Oct 2026
How and when it might arrive
When could superintelligence happen?
What forecasters actually say about the arrival of AGI and superintelligence, why they disagree, and why the US administration's "Super Intelligence" label is not a forecast at all.
Last reviewed 23 Sept 2026What is an intelligence explosion?
The idea that AI improving AI could compress years of progress into months: where it comes from, why automated AI research revived it in 2026, the bottlenecks and the objections. A hypothesis, not a forecast.
Last reviewed 3 Oct 2026What is recursive self-improvement?
AI helping to build better AI: what the term means, how much of AI development AI already does, why compute, hardware, data and experiments limit the loop, and why it does not automatically mean runaway growth.
Last reviewed 3 Oct 2026What is AI takeoff? Fast and slow scenarios
How quickly AI might go from human-level to far beyond it: fast and slow takeoff, continuous and discontinuous progress, the arguments on each side, and why the shape matters more than the date.
Last reviewed 3 Oct 2026
What is the technological singularity?
Where the idea of a technological singularity comes from (von Neumann, Good, Vinge, Kurzweil), what it claims, and why it is not the same thing as AGI, superintelligence or an intelligence explosion.
Last reviewed 3 Oct 2026
What it could do, and the risks
What could superintelligence actually do?
Science, mathematics, software, medicine, materials, persuasion, strategy, the economy and AI research itself: what today's AI has demonstrated in each, kept separate from what is argued about a hypothetical superintelligence.
Last reviewed 3 Oct 2026What are the risks of superintelligence?
Misuse, accidents, misalignment, concentration of power, cyber and military risks, economic disruption and loss of control, each separated into what is observed in today's AI and what is argued about hypothetical superintelligence.
Last reviewed 3 Oct 2026What is the AI alignment problem?
Why it is hard to make AI pursue the goals its designers intend: specification gaming, outer and inner alignment, laboratory evidence of deceptive behaviour, scalable oversight, and why researchers disagree about how hard it is.
Last reviewed 3 Oct 2026
Can superintelligence be controlled?
The control problem in plain English: corrigibility and the off switch, sandboxing and the 2026 escape incidents, monitoring, interpretability, compute limits and governance, and why none has been shown to work on a system smarter than its overseers.
Last reviewed 3 Oct 2026Frontier AI safety frameworks: how AI companies decide a model is too dangerous
What a frontier AI safety framework is, how capability thresholds, dangerous-capability tests and safeguards work, when a company says it would slow down, what is voluntary and what is law, and what these documents cannot tell you.
Last reviewed 28 Oct 2026
Who is building it
Who is building superintelligence?
Which organisations say they are aiming at AGI or superintelligence, what each has said in its own words, and why the US administration's "Super Intelligence" label covers a much wider field.
Last reviewed 23 Sept 2026OpenAI and AGI: what is it trying to build?
What OpenAI says it is building, how it defines AGI, why its leaders now talk about superintelligence, what its models have demonstrated, and how its safety approach and the September 2026 pause fit in.
Last reviewed 28 Oct 2026Google DeepMind and AGI: what is it trying to build?
How Google DeepMind defines AGI, how its researchers propose to measure it, what they have written about the path to superintelligence, what it has demonstrated, and its Frontier Safety Framework.
Last reviewed 28 Oct 2026Anthropic and advanced AI: what is it trying to build?
Why Anthropic talks about "powerful AI" rather than AGI, what its Claude models and Responsible Scaling Policy involve, what it reports about AI doing its own research, and what independent testing found in 2026.
Last reviewed 28 Oct 2026Meta Superintelligence Labs: what does Meta mean by superintelligence?
What Meta Superintelligence Labs is, what Meta means by "personal superintelligence", how that differs from the research meaning of the word, and what its Muse models and safety framework show.
Last reviewed 28 Oct 2026Safe Superintelligence Inc.: what is it trying to build?
What is publicly known about Safe Superintelligence Inc., Ilya Sutskever's company with one stated goal and one product: a safe superintelligence. Its statements, funding and the large gaps in the record.
Last reviewed 3 Oct 2026xAI (now SpaceXAI): what is it trying to build?
What xAI, now part of SpaceX as SpaceXAI, says it is for, its Grok models and Colossus computing sites, how its safety frameworks treat superintelligence, and the UK investigations into Grok.
Last reviewed 28 Oct 2026China's frontier AI labs: who is building advanced AI?
DeepSeek, Moonshot AI, Zhipu AI, Alibaba's Qwen, ByteDance Seed and MiniMax: what China's leading AI developers say they are building, what independent testing shows, and what cannot be known.
Last reviewed 28 Oct 2026
Reference
Superintelligence glossary: the words you will meet, in plain English
Short, accurate definitions of the terms used in discussions of AI, AGI and superintelligence, from alignment to context window, with the jargon kept to a minimum.
Last reviewed 23 Sept 2026