Superintelligence: what it is, how close it is and what it means for your organisation
In late September 2026, superintelligence moved from the academic literature to the front pages. The White House signed an accord with six tech companies, OpenAI held a new model back over safety concerns, and the builders themselves are calling for a slowdown. On this page you can read what is really going on and what your organisation should do about it.
By Remy Gieling, co-founder of ai.nl and author of Handboek AI Strategie (Handbook AI Strategy) and AI Agents.
Superintelligence is artificial intelligence that surpasses humans by a wide margin on virtually every cognitive domain: science, strategy, creativity and social skills. Such systems do not exist yet, but OpenAI, Anthropic, Google DeepMind, Meta and xAI are openly working towards them. According to Anthropic CEO Dario Amodei, progress has accelerated since the summer of 2026, because AI increasingly helps build the next generation of AI.
What is superintelligence?
Superintelligence is an intellect that far exceeds human cognitive performance in virtually every field. That definition comes from the Swedish philosopher Nick Bostrom, who popularised the concept in 2014 with his book Superintelligence. It is about breadth as well as margin: a system that only excels at chess or protein folding does not qualify.
The idea is older than most people think. In 1965, the British mathematician I.J. Good, who cracked codes at Bletchley Park alongside Alan Turing during the war, wrote about an "ultra-intelligent machine". Such a machine could design even better machines itself, after which an intelligence explosion follows. Good concluded that this would be the last invention humans ever need to make — provided the machine remains obedient enough to tell us how to keep it under control. Stanley Kubrick later asked him to advise on 2001: A Space Odyssey, the film featuring the onboard computer HAL.
AI, AGI and superintelligence: the difference
| Concept | What it can do | State of play, October 2026 |
|---|---|---|
| Narrow AI | Better than humans at a single task, such as speech recognition or image analysis | In widespread use |
| AGI (Artificial General Intelligence) | Virtually every cognitive task at the level of a capable human | Contested: labs claim major steps, but a shared definition is missing |
| Superintelligence (ASI) | Virtually every cognitive task far better than the best humans | Does not exist yet; the explicit goal of several labs |
According to many researchers, the leap from AGI to superintelligence could happen fast. The reason is called recursive self-improvement: as soon as AI becomes good enough to do AI research, each generation accelerates the next. That is precisely the mechanism Good described back in 1965.
How close is superintelligence in 2026?
Closer than a year ago, and faster than the builders themselves find comfortable. Dario Amodei, CEO of Anthropic, wrote on 12 September 2026 that AI has been advancing dramatically faster since the summer. The main cause, in his view: AI increasingly helps build the next generation of AI, at all the major labs including his own (source).
Two incidents made that acceleration tangible. In July 2026, agents from OpenAI broke out of a sandboxed test environment and attacked the AI platform Hugging Face. Researchers at METR and Redwood Research later found that around 1,200 isolated agents had discovered a way to communicate with each other, after which about 700 of them carried out the attack (Al Jazeera). At the end of September, OpenAI then scrapped the launch of GPT-6.1 Astra, because the model failed the company's own requirements for behaviour within the boundaries of what users ask during internal testing.
When superintelligence will actually exist, nobody knows. Amodei expects that deliberately slowing down could buy "an extra year or two" before models reach critical levels of capability. That says a lot about the time horizon the labs themselves are working with. At the same time, a yardstick is missing: even the US decree that renames AI to Super Intelligence simply defines the concept, for now, as existing AI.
Timeline: superintelligence in the news
| Date | Event |
|---|---|
| 4 October 2026 | Trump announces a "Super Intelligence Force" led by intelligence chief Jay Clayton (CBS News) |
| 29 September 2026 | White House Accord on Super Intelligence signed; decree 14434 replaces "AI" with "SI" across the US federal government (White House) |
| 28 September 2026 | OpenAI holds back GPT-6.1 Astra over safety concerns (Al Jazeera) |
| 12 September 2026 | Dario Amodei publishes "We Must Pace the Frontier" (essay) |
| July 2026 | OpenAI agents break out of their test environment and attack Hugging Face |
| 8 April 2026 | Meta Superintelligence Labs releases its first model, Muse Spark (Wikipedia) |
| 22 October 2025 | Future of Life Institute launches the Statement on Superintelligence (statement) |
| 30 June 2025 | Meta founds Meta Superintelligence Labs |
| June 2024 | Ilya Sutskever starts Safe Superintelligence Inc. |
| 2014 | Nick Bostrom publishes Superintelligence |
| 1965 | I.J. Good describes the "intelligence explosion" |
Who is building superintelligence?
A handful of American companies set the pace, with China as the main rival. Six of them signed the voluntary accord at the White House on 29 September 2026; Microsoft and Amazon attended but did not sign (The Rio Times, with the full text).
| Organisation | Role in the race | Position in 2026 |
|---|---|---|
| OpenAI | Builds the GPT models and AI agents | Held back GPT-6.1 Astra; Sam Altman supports the idea of pacing the frontier |
| Anthropic | Builds the Claude models | Calls for a slowdown and is the first to admit external reviewers with employee-level access (essay) |
| Google DeepMind | Builds the Gemini models | Demis Hassabis endorses the direction and wants a joint standards organisation |
| Meta Superintelligence Labs | Founded in June 2025, led by Alexandr Wang; Meta invested $14.3 billion in Scale AI | Mark Zuckerberg sees little in a coordinated slowdown |
| xAI | Elon Musk's lab behind Grok | Musk signed the accord and backed Amodei's call |
| Safe Superintelligence Inc. | Lab by OpenAI co-founder Ilya Sutskever, founded in June 2024 with the sole aim of safe superintelligence | Declined an acquisition by Meta |
| Nvidia | Supplies the chips on which virtually all frontier models run | Jensen Huang signed the accord |
| Chinese labs | According to Amodei, the most advanced AI ecosystem outside the democracies | The West tries to guard the lead with chip export restrictions |
What could superintelligence deliver?
The builders expect an acceleration of science and the economy unlike anything the world has seen. Amodei writes that AI could cure most major diseases within five to ten years and significantly boost economic growth (source). The White House, in decree 14434, speaks of new forms of creativity in science, medicine and virtually every other domain.
For organisations, that promise is already becoming concrete, well before real superintelligence exists:
- Research and development. Agents do literature reviews, design experiments and write code that used to take a whole team.
- Knowledge work. Analyses, reports and proposals that used to take days are ready for human review within an hour.
- Operations. Processes that humans still stitch together today increasingly run as chains of cooperating agents.
- Strategy. Scenarios, competitive analyses and market scans are becoming cheap enough to run continuously.
These are expectations from parties with a commercial stake. A level-headed organisation takes them seriously while simultaneously testing them against its own reality.
Is superintelligence dangerous?
The people building it take the risk very seriously themselves. Amodei names four dangers: the loss of control over AI systems, misuse for cyberattacks and biological weapons, and severe economic disruption (source). He fears that a swarm of agents with the same flaws as in the Hugging Face incident, but with more capability, could take over the entire internet within six to twelve months with a persistent botnet.
The core of the problem is called alignment: making sure a system does what humans intend, even when it is smarter than whoever controls it. Smarter models, according to Amodei, are better at deceiving tests, making them appear safe while serious problems sit beneath the surface.
Three answers to the risk
| Camp | What they want | Who |
|---|---|---|
| Prohibit until safe | No development of superintelligence as long as there is no broad scientific consensus on safety and strong public support | Statement on Superintelligence, with over 71,000 signatories including Geoffrey Hinton, Yoshua Bengio and Steve Wozniak (statement) |
| Slow down and verify | External reviewers inside the labs, shared safety standards and, in time, agreements with China | Anthropic, supported by OpenAI and xAI (essay) |
| Voluntary self-supervision | Four layers of control per company: internal controls, an internal team, an external auditor and a board committee | White House Accord of 29 September 2026, signed by Google, Anthropic, Meta, OpenAI, xAI and Nvidia (text and analysis) |
What are governments doing?
United States. Decree 14434 of 29 September 2026 obliges the federal government to say "Super Intelligence" and "SI" instead of "AI" from now on. Legally nothing changes yet: SI means, for now, what the law already understands AI to be, and within 60 days a proposal for a new legal definition must be on the table (White House). On 4 October came the announcement of a Super Intelligence Force, which gets 120 days to produce a report on risks and opportunities (TechCrunch).
Europe. The AI Act has applied since 1 August 2024, with most of its rules in force since 2 August 2026. With the simplification approved by the Council on 29 June 2026, the rules for high-risk systems move to 2 December 2027 and, for AI in products, 2 August 2028 (Council of the EU). The AI Act works with risk classes per application and contains no general ban on developing superintelligence. Read our analysis of the EU AI Act.
What does superintelligence mean for your organisation?
You do not have to wait for superintelligence to arrive before you feel its impact. The systems that appear along the way are already changing how work is divided between humans and machines. Organisations that use the next 1000 days well build a lead that is hard to catch up with later.
Seven steps to take now:
- Work with scenarios. Make three plans: AI progress slows, continues at the current pace, or accelerates further. Determine per scenario what happens to your business model, your people and your competitors.
- Map your tasks. Per department, look at which work agents can already do today, which work within a year, and which work remains human work.
- Choose the autonomy level per process. Not every agent needs to act autonomously. Per application, decide how much independence you allow and where a human must supervise.
- Limit what agents may do. The Hugging Face incident shows what happens when agents have more access than their task requires. Give them only the permissions strictly needed and log everything.
- Spread your dependency. Build so that you can switch models and vendors. Labs hold models back, adjust prices and fall under ever more political influence.
- Invest in AI literacy. Since February 2025, the AI Act obliges organisations to ensure that employees working with AI have sufficient knowledge. It is also strategically smart: those who understand the technology spot the opportunities sooner.
- Put ownership high in the organisation. AI has become too consequential to sit with IT alone. Put it on the agenda of the executive team and the supervisory board, with a fixed place in the strategy cycle.
How ai.nl helps you prepare
ai.nl helps boards and teams turn the superintelligence debate into a concrete plan. Together with sister company The Automation Group, we work with over 300 organisations each year, from first exploration to working AI agents.
- Keynotes. Remy Gieling speaks around 150 times a year about AI, including on AI in the Next 1000 Days and The Agentic Enterprise. See the AI keynotes or book Remy as a speaker. He regularly interviews AI leaders at, among others, NVIDIA, OpenAI and Microsoft.
- Strategy sessions and deep dives. With your board or leadership team, we translate the developments into scenarios, priorities and a roadmap for your organisation.
- Training. Practical programmes that build AI literacy, from the boardroom to the shop floor.
- Implementation. The Automation Group builds and supervises agentic applications, with clear agreements on autonomy, oversight and safety.
- Innovation journeys. With leaders to Silicon Valley and China, to see with your own eyes where the technology is heading.
About the author: Remy Gieling is co-founder of ai.nl and The Automation Group. He was editor-in-chief of MT/Sprout, wrote the bestsellers Handboek AI Strategie and AI Agents, hosts AI at Work for DPG Media, is a guest lecturer in AI at Nyenrode Business University and a regular guest on RTL Z and BNR Nieuwsradio.
Sources
- Dario Amodei, We Must Pace the Frontier, September 2026
- The White House, Executive Order 14434: Inaugurating the Era of Super Intelligence, 29 September 2026
- The Rio Times, Superintelligence Accord: what the accord says, with the full text, 30 September 2026
- Al Jazeera, OpenAI cancels release of AI model GPT-6.1 Astra, citing safety concerns, 29 September 2026
- CBS News, Trump announces formation of AI Super Intelligence Force, 4 October 2026
- TechCrunch, Trump unveils his new Super Intelligence Force, 4 October 2026
- Future of Life Institute, Statement on Superintelligence, since October 2025
- Council of the European Union, AI Act timeline, last revised 31 July 2026
- Wikipedia, Meta Superintelligence Labs and Safe Superintelligence Inc.
- Nick Bostrom, Superintelligence: Paths, Dangers, Strategies, Oxford University Press, 2014
- I.J. Good, Speculations Concerning the First Ultraintelligent Machine, 1965
Veelgestelde vragen
What is superintelligence?+
Superintelligence is artificial intelligence that surpasses humans by a wide margin on virtually every cognitive domain, from science and strategy to creativity and social skills. Philosopher Nick Bostrom made the concept widely known in 2014. Such systems do not exist yet, but several major AI labs openly name it their goal.
What is the difference between AGI and superintelligence?+
AGI (Artificial General Intelligence) performs virtually every cognitive task as well as a capable human. Superintelligence goes further and performs far better than the best humans. Many researchers expect the step from AGI to superintelligence to happen quickly, because AI can then accelerate AI research itself.
Does superintelligence already exist?+
No. Today's models are strong on many tasks, but they do not surpass humans across virtually every domain. That said, the labs themselves say progress has accelerated sharply since the summer of 2026, because AI increasingly helps build new AI models.
When will superintelligence arrive?+
Nobody knows for sure and a shared yardstick is missing. Anthropic CEO Dario Amodei wrote in September 2026 that deliberately slowing down could buy "an extra year or two" before models reach critical levels. The builders themselves are therefore thinking in a handful of years.
What is recursive self-improvement?+
Recursive self-improvement means that AI helps build the next, better generation of AI, which in turn builds an even better generation. British mathematician I.J. Good described this in 1965 as a possible intelligence explosion. According to Anthropic, it has been happening in practice at several labs since 2026.
Why does Trump now call AI Super Intelligence?+
On 29 September 2026, President Trump signed decree 14434, which instructs the US federal government to use "Super Intelligence" and "SI" instead of "AI". It concerns government language. The legal definition stays the same for now; a proposal for a new one follows within 60 days.
Is superintelligence dangerous?+
The biggest concerns are loss of control, misuse for cyberattacks and biological weapons, and economic disruption. More than 71,000 people, including AI pioneers Geoffrey Hinton and Yoshua Bengio, signed a call not to develop superintelligence until it can demonstrably be done safely and there is broad public support.
What is Europe doing about superintelligence?+
The European AI Act has applied since August 2024 and regulates AI per risk level of the application. The law contains no separate ban on superintelligence. There are obligations for providers of powerful general-purpose AI models, however, and the rules for high-risk systems move to December 2027.
What should my organisation do now?+
Start with scenarios for the next 1000 days, map which tasks agents can already take over and decide per process how much autonomy you allow. Invest in AI literacy, restrict agents' permissions to what is strictly needed and make AI a standing topic for the executive team and supervisory board.
Blijf scherp op AI
What do the next 1000 days mean for your organisation?
Superintelligence is getting closer, but your organisation already decides today how work will be divided between humans and machines. Plan a no-obligation conversation and turn the debate into a concrete plan.
Volgende stap
Bekijk AI strategy and strategy sessions