Originally published by PwC on 20 May 2026
The debate about the future of Artificial Intelligence has increasingly been shaped by concepts such as Artificial General Intelligence (AGI), Technological Singularity and Superintelligence, often invoked to describe scenarios of profound transformation, but not always clearly understood.
In this article, Ricardo Pacheco Almeida, AI Tech Lead at PwC Portugal, traces these ideas back to their origins and charts the path towards achieving superhuman intelligence.
Did you know that AGI, Technological Singularity and Superintelligence describe distinct perspectives on the possible future of AI?
- AGI refers to systems capable of performing a wide range of cognitive tasks at a human level of performance;
- Technological Singularity denotes the point at which this capability evolves autonomously and at an accelerating pace; and
- Superintelligence represents a stage at which machines systematically outperform humans in virtually all intellectual activities.
For leaders and organisations, understanding these distinctions is more than an academic exercise. How AGI is defined shapes how leaders and organisations interpret its future potential and the associated impacts in areas such as regulation, investment, risk management and strategic planning.
Thus, this article does not seek to anticipate the potential impacts on society, the environment, the economy and politics arising from the emergence of these phenomena, but rather to focus on clarifying concepts — the foundation of any future discussion.
From the beginnings of statistical AI to GenAI
To understand AGI, Technological Singularity and Superintelligence, it is worth looking back to a time when AI did not yet dominate public debate. When I was studying the Economics of Innovation, years before the emergence of commercial models such as ChatGPT, Artificial Intelligence was not discussed as frequently in this field of study as it is today.
The concept of Artificial Intelligence (AI) is already several decades old. Despite multiple moments of “hype” over the years, for example when Deep Blue defeated world chess champion Garry Kasparov in 1997, AI was still largely seen as a set of fundamentally statistical techniques for prediction, classification and pattern recognition, in other words, typical use cases for “traditional” Machine Learning.
There was already speculation about how technology might give rise to intelligence superior to that of humans. Two schools of thought emerged with radically different visions: one in Europe, the other in the United States.
Soft-Singularity: the Global Brain
In Belgium, from the 1990s onwards, cyberneticist Francis Heylighen developed the concept of Emergence and the Global Brain (or Soft-Singularity): a passive event in which consciousness would emerge from the enormous volume of data and connections exchanged among billions of humans and machines. This phenomenon would be identical to the way human neurons work: individually, they have no intelligent capabilities, but together they form a brain. The whole goes beyond the sum of its parts.
In other words, the internet was expected to “wake up” by itself, with a consciousness, based on the amount of information it contained. It would therefore be a collective, distributed meta-intelligence, of which connected humans themselves would be part, and its emergence would depend solely on the number of existing connections.
Hard-Singularity: the Technological Singularity
Whereas the European vision relied on passive emergence, with intelligence arising naturally from the connections between humans and machines, the American vision proposed something more deliberate and potentially more disruptive. In Silicon Valley, the concept of the Technological Singularity (or Hard-Singularity) emerged, popularised by futurists such as Ray Kurzweil (e.g. in the books The Singularity is Near, 2005, and The Singularity is Nearer, 2024).
The term “Singularity” is a metaphor borrowed from physics and mathematics. It refers strictly to the point at which a variable (such as density in a black hole) tends towards infinity and the apparent laws of physics cease to make sense. This event would mark a turning point in history at which technological evolution becomes so dizzyingly fast that human forecasting models collapse, making it impossible to anticipate the future.
The central concept here is the “Intelligence Explosion”, originally described by mathematician and statistician I.J. Good: the moment when we create an AGI capable of designing an improved version of itself. This process would initiate a cycle of recursive self-improvement: version 1.0 creates 2.0; 2.0, being more intelligent, creates 3.0 more quickly, and so on. This cycle would repeat thousands of times at electronic speeds, resulting in a Superintelligence (ASI) that would surpass human comprehension in days or even hours.
Decades after I.J. Good and Ray Kurzweil, this idea not only remains alive but has gained urgency. It is the predominant idea behind the “AI 2027” report, published in 2025 by Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland and Romeo Dean, which sets out to predict what the year 2027 will be like, assuming the development of an AGI capable of self-improvement.
What is AGI?
All scenarios associated with the Technological Singularity and Superintelligence start from one assumption: the emergence of AGI. But what, specifically, is AGI? And how far along the path are we?
Defining AGI has proved complex, with scientific consensus fragmenting into distinct corporate definitions. As of today, OpenAI defines it in terms of economic impact (the replacement of human labour), Anthropic focuses on risk and safety levels (ASL), and Meta (through Yann LeCun) rejects the term, arguing that current models lack “world models” and “physical” understanding.
The perspective I tend to adopt is the one proposed by the Google DeepMind research laboratory, whose central premise is “focus on capabilities, not processes”. This means that, for a system to be considered AGI, it need only perform tasks with equivalent competence; it does not need to possess “sentience” or mimic human thought. Essentially, whatever the task, an AGI should be able to carry it out — from creating an accurate presentation on Greek mythology to making a call to book the most suitable restaurant in town.
“The different perspectives on AGI, Singularity and Superintelligence remain relevant as different angles on the future of Generative Artificial Intelligence. They all serve as tools and sources of inspiration for defining long-term strategies, particularly in sectors where Generative AI has the greatest impact.”
Ricardo Pacheco Almeida, AI Tech Lead at PwC Portugal
The five levels of AGI
To clarify how AGI may evolve over time, Google DeepMind proposes a model that describes it as an evolutionary scale, rather than a binary breaking point. It is based on two axes that, in turn, result in five classification levels:
- Performance (depth of competence); and
- Generality (breadth of tasks).
Classification matrix for the evolution of AGI
Level 1 (Emerging AGI)
Systems that possess generality, but whose overall performance is comparable only to that of an unskilled human. This is where DeepMind places current frontier models, such as the latest from OpenAI, Anthropic and Google.
Level 2 (Competent AGI)
The threshold at which the system matches the performance of 50% of skilled adults across a wide range of tasks.
Levels 3 and 4 (Expert and Virtuoso)
Where capability scales to the 90th and 99th percentiles of skilled adults, respectively.
Level 5 (Superhuman AGI)
Systems that outperform 100% of humans on all tasks.
We can consider that, upon reaching level 5, we will be facing a Superintelligence, a virtual entity with a level of performance superior to that of all humans on all tasks, in general. Access to a technology with this potential will have a very profound impact on everything we know.
Multi-agent AGI in the future?
The evolution of AGI may not involve a single, increasingly powerful model, but systems composed of multiple specialised agents acting in coordination. The question now is: how do we get there? Through a single, increasingly powerful model, or by a different route?
With the recent emergence of GenAI agents, which can be defined as a software program that uses GenAI to make decisions and use tools, DeepMind has been suggesting that AGI could arise naturally from systems of agents with multiple specialisations and tools at their disposal, rather than from a highly advanced monolithic brain. This increasingly popular perspective has been dubbed “multi-agent AGI or collective intelligence.”
Interestingly, this vision of intelligence emerging from collaboration among multiple specialised agents recalls Heylighen’s Global Brain, with one fundamental difference: it does not depend on passive emergence from the internet, but on intentionally designed systems.
In a debate that oscillates between enthusiasm and apprehension, clarifying these concepts is essential if leaders, organisations and decision-makers are to place the evolution of AI in an informed and responsible context. This awareness should guide a strategic, prudent and forward-looking approach to AI throughout this century.