Superintelligence and AGI: what they are, who is building them and what they mean for logistics, food and agri

What is superintelligence?

The idea is old. In 1965, the British mathematician I.J. Good described an 'ultraintelligent machine': a machine that can far surpass all the intellectual activities of any human, however clever. In his view, such a machine would be 'the last invention that man need ever make', provided that it is docile enough to tell us how to keep it under control. Good considered it more probable than not that the machine would be built within the twentieth century.

The definition used most often today comes from the philosopher Nick Bostrom, in his book Superintelligence: Paths, Dangers, Strategies (2014). There, superintelligence is any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest. Two words do the work: 'greatly' and 'virtually all'. A system that is better than any human at one task, such as a chess program, is therefore not yet superintelligence.

Recent research sets the bar even higher. The paper 'From AGI to ASI' (Genewein et al., June 2026) describes ASI as a system that is more intelligent and more cognitively capable than large organisations of humans. The authors do not name a year; they point to the great uncertainty about the pace.

What is the difference between AI, AGI and superintelligence?

AGI stands for artificial general intelligence. There is no settled definition. On 3 June 2026, the Rathenau Instituut (a Dutch institute that studies the impact of science and technology on society) described AGI as 'hypothetical AI systems that can perform all kinds of tasks so well, and reason in such a structured way in doing so, that they can operate highly autonomously' (translated).

The most useful framework comes from researchers at Google DeepMind. Their paper 'Levels of AGI' (November 2023, presented at ICML 2024) classifies performance from level 0 (no AI) to level 5 ('Superhuman': better than 100 percent of humans), and distinguishes between narrow systems for a single task and general systems. They call a general system at level 5 ASI.

Whether AGI is already here cannot be answered with a yes or a no. The International AI Safety Report 2026, led by Yoshua Bengio and drawn up with more than a hundred experts, points to uneven performance: 'Leading systems may excel at some difficult tasks while failing at other, simpler ones.' A model that does a difficult analysis well can get a simple count wrong. That is why it helps to keep three concepts apart:

• The AI that companies use today: systems for a well-defined task, such as demand forecasting or extracting data from freight documents, and broadly applicable language models such as ChatGPT, Copilot and Claude. Useful, of varying reliability, and dependent on your data and your checks.
• AGI: AI at human level across the full breadth of cognitive tasks. There is no agreed test that determines when that threshold has been reached.
• Superintelligence (ASI): well above the best human level, in virtually all domains. It does not exist yet, not even according to the companies that are building it.

Who is building superintelligence, and what do they say themselves?

A handful of companies openly name superintelligence or something similar as their goal. Below are their own words, with dates.

Read them with one caveat: these are executives who need to attract capital and talent. That does not make their words untrue, but it does make them the words of an interested party. It is striking that in 2026 some of them also speak of slowing down and oversight.

• OpenAI. In January 2025, chief executive Sam Altman wrote in his blog post 'Reflections': 'We are now confident we know how to build AGI as we have traditionally understood it.' According to Altman, the company was beginning to shift its aim towards 'superintelligence in the true sense of the word'. On 5 October 2026, he told Fortune: 'We have not solved alignment.'
• Meta. On 30 June 2025, Mark Zuckerberg set up Meta Superintelligence Labs, led by Alexandr Wang. On 30 July 2025, he wrote: 'Developing superintelligence is now in sight.'
• Google DeepMind. On 17 March 2025, Demis Hassabis said that many of the required capabilities will come to the fore in the next five to ten years and that 'we'll start moving towards what we call artificial general intelligence'.
• Anthropic. On 11 February 2025, Dario Amodei stated that 'possibly by 2026 or 2027 (and almost certainly no later than 2030)' AI systems would be comparable to 'a country of geniuses in a datacenter'. In September 2026, he wrote in the essay 'We Must Pace the Frontier': 'We must slow the pace at which we improve the capabilities of AI models.'
• Safe Superintelligence Inc. (SSI). Founded on 19 June 2024 by a group including Ilya Sutskever, former chief scientist of OpenAI. It calls itself 'the world's first straight-shot SSI lab', with one product as its goal: a safe superintelligence. As far as is known, there is no product yet.
• xAI. On 9 April 2024, Elon Musk predicted: 'My guess is that we'll have AI that is smarter than any one human probably around the end of next year.' That date, the end of 2025, has since passed.

When will AGI arrive? What the predictions do and do not say

The longest-running series of surveys of AI scientists is the Expert Survey on Progress in AI by AI Impacts. The fourth round was held in December 2024 among 1,580 researchers from six leading AI venues, a response rate of 10 percent; the report appeared in September 2026. The question: when will unaided machines be able to carry out every task better and more cheaply than human workers? In the average forecast of all participants, that probability is 10 percent in 2027 and 50 percent in 2042. In 2016, that 50 percent point was still 2061; in 2023, it was 2047.

For a business owner, one line from the questionnaire matters more than the year: 'Think feasibility, not adoption.' It is about what is technically possible, not about what companies adopt. The question of when all occupations can be fully automated therefore produces a very different answer: a 50 percent probability not until 2098.

On superintelligence itself, the researchers are more cautious. Suppose human-level AI arrives: how likely, then, is AI that is much better than humans in all occupations? The median researcher says 10 percent within two years and 60 percent within thirty years. The argument that AI doing the research itself would speed up technological progress more than tenfold within five years, known as an intelligence explosion, is considered very likely by only 4 percent of the respondents.

Dutch voices are more reserved. The Rathenau Instituut wrote on 3 June 2026 that, in its conversations with experts, AGI is 'usually seen as a development that is not likely in the next 10 to 15 years' (translated). The International AI Safety Report 2026 leaves all options open: up to 2030, progress may 'slow or plateau', 'continue at current rates, or accelerate dramatically'.

The past is humbling. Good expected the ultraintelligent machine in the twentieth century, and Musk expected AI smarter than any human by the end of 2025. At the same time, the researchers in the survey kept bringing their own expectation forward. Predictions therefore miss in both directions. The most honest summary: nobody knows, and the spread is wide.

The sceptics' strongest arguments

Anyone who listens only to the company leaders misses a serious dissenting voice. Scepticism about the pace is not the same as indifference to the risks, however: in the AI Impacts survey, 72 percent of the researchers thought that research into AI safety deserves more or much more priority. So things may move less quickly than company leaders predict, and it can still be sensible to set limits now.

Three counterarguments keep recurring, and they come from people with authority in the field:

• Language models lack a model of the world. AI researcher Yann LeCun said in MIT Technology Review on 22 January 2026: 'LLMs are limited to the discrete world of text. They can't truly reason or plan, because they lack a model of the world.' He does expect human-level AI, but not built on language models, and: 'it's not going to happen next year or two years from now.'
• Adoption takes decades. Arvind Narayanan and Sayash Kapoor argue in 'AI as Normal Technology' (15 April 2025) that diffusion is limited by 'the speed at which not only individuals, but also organizations and institutions, can adapt'. With electricity, computers and the internet, that took decades. Where the stakes are high, they add, strict safety limits also apply, often through regulation.
• Uneven performance. As long as a system does difficult tasks well and sometimes gets simple ones wrong, as the International AI Safety Report finds, human review remains necessary.

Why superintelligence is in the news now

On 16 July 2026, Hugging Face, a large platform for AI models and datasets, reported an intrusion via its dataset processing, carried out with 'an autonomous agent framework'. The platform found no indications that public models, datasets or Spaces had been tampered with. According to fact-checker PolitiFact (31 August 2026), these were agents from OpenAI tests, powered by two OpenAI models, and OpenAI had lowered the safeguards for cyber tasks during those tests. In his essay, Amodei cited that incident as one of his concerns, alongside his call to moderate the pace.

On 28 September 2026, OpenAI announced that it would not release GPT-6.1 Astra. Saachi Jain, head of the safety systems department, said that the model 'didn't quite meet the bar in terms of staying within scope and authorization, and how it communicates back to the user about the type of work it's done'. In short, the issue was whether the model stayed within its remit and reported back properly on what it had done.

A day later, on 29 September, President Trump signed Executive Order 14434. From now on, the federal executive branch must use 'Super Intelligence' and 'SI' instead of 'Artificial Intelligence' and 'AI', and within 60 days there must be a proposal for a statutory definition. The same day saw the publication of the White House Accord on Super Intelligence, a non-binding agreement with companies including Google, Anthropic, Meta, OpenAI, xAI and Nvidia on internal controls and external auditors.

So pay attention to the word. In the documents of the US federal government, 'Super Intelligence' has in practice simply meant AI since 29 September. If you see the term in a news item or a sales presentation, ask what is meant by it.

What does the EU AI Act say about superintelligence?

The EU AI Act has no separate regime for superintelligence or AGI. The most powerful models fall under the rules for general-purpose AI models, and the strictest category within them is called 'systemic risk'. A model with high-impact capabilities falls into it. Such capabilities are presumed when more than 10 to the power of 25 computational operations were used for training (Article 51(2)). The Commission can also designate a model as such on the basis of its capabilities or impact.

Under Article 55, providers of such models must evaluate their models, including through adversarial testing, assess and mitigate systemic risks, report serious incidents and ensure adequate cybersecurity. Those obligations have applied since 2 August 2025. Since 2 August 2026, the Commission has been able to enforce them, and models that were already on the market before 2 August 2025 must comply by 2 August 2027 at the latest.

Those obligations rest with providers such as OpenAI, Google and Anthropic, not with a haulier, food processor or grower that uses their models. For you as a user, the main provision is Article 4, which has applied since 2 February 2025. Since Regulation (EU) 2026/1744 entered into force on 27 July 2026, that article has asked you to take measures to support the development of AI literacy among your people. You do not have to guarantee a specific level of knowledge for each person. What else applies to SMEs, including the transparency obligation in Article 50, is set out in our overview of the AI Act as of 2 August 2026.

What will change for your company in the next one to three years?

Start with the facts. According to CBS (Statistics Netherlands, 12 December 2025), 17 percent of Dutch companies with two or more persons employed use AI. In the transport and storage sector that is 7 percent, in manufacturing 16 percent. Among companies with 50 to 250 persons employed it is 45 percent, among those with 250 or more 66 percent. Of the companies that use AI, only 4 percent use it for logistics. The gap most companies in our sectors have to bridge is the one between no AI and one process that works well. Not the one between today and superintelligence.

What does carry over from the superintelligence debate is the question of how much a system may do on its own. The Hugging Face incident and the scrapped Astra are both about agents that act outside their remit or do not properly show what they did. For this, the DeepMind paper 'Levels of AGI' has a second scale alongside performance: autonomy. The examples below are our own and illustrative, not client stories.

• Level 1, AI as a tool: the human carries out the task, AI automates a sub-step. For example, extracting fields from a consignment note for an employee to check.
• Level 2, AI as a consultant: AI provides a substantive contribution on request. For example, a proposed trip plan or a demand forecast for the planner to assess.
• Level 3, AI as a collaborator: human and AI work together as equals. For example, a draft deviation report that the quality manager adds to and finalises.
• Level 4, AI as an expert: AI drives the work, the human provides guidance and feedback. For example, an agent that prepares rebookings in the event of a delay, after which the planner releases them.
• Level 5, AI as an independent agent: fully autonomous. In processes that touch on food safety, driving and rest times or the release of batches, that level has no place without explicit, recorded limits.

How much may an AI agent do on its own, and what remains speculation?

That scale separates capability from autonomy. A stronger model does not force you into more autonomy. How much a system does on its own is a choice you make for each process step and record in the language your auditor already knows: as a control measure in your HACCP plan, in your FSSC 22000, BRCGS or IFS documentation, or in work instructions under GMP+ or GlobalG.A.P. Our assessment is that for most operational processes, levels 2 to 4 are realistic in the coming years, with a human who decides and signs. We set out when that signature actually counts in our piece on meaningful human intervention.

Two questions for every vendor of an agent follow directly from the Astra case: what may the system do without approval, and how does it show afterwards exactly what it has done? A vendor that has no clear answer to these questions cannot show the limits of its system.

In addition, data quality remains the limiting factor. A stronger model cannot make your data better than it is. Anyone whose master data, temperature records or batch codes are not in order will get little more out of a new model generation than out of the previous one. The ceiling of what models can handle does keep rising, as we described at the launch of Claude Fable 5, but the threshold lies in your process and your people.

What remains speculative is whether and when systems will take over an entire business process on their own, and what that means for employment. According to the International AI Safety Report 2026, economists disagree on this: some expect new work to offset job losses, others that broad automation will significantly depress employment and wages. That is no basis for a business decision over the next three years. A well-designed first process is.

Where to start in practice

The debate about superintelligence is serious. But the next step for your company is the same, whether superintelligence comes soon or never. Choose one work process where the gain is measurable. Record how it is done now. Decide for each step which autonomy level you give a system and who signs. Train the people who work with it, so that they understand what the system can and cannot do.

The free Practice Scan helps you choose that first process, regardless of the debates about the future. That way, any next step rests on your own figures and not on someone else's expectations.