For decades, the technological singularity lived mainly in science fiction, futurist essays, and a few corners of computer research.
Today, it is back at the center of the conversation.
We hear that “the singularity is near,” that artificial intelligence could cause it, or even that the process has already begun. The problem is that the word is often used without a clear idea of what it describes.
The singularity is not simply a very powerful AI. Nor is it a synonym for artificial general intelligence.
It is the hypothesis that technological progress could reach such a profound point of acceleration that our ability to anticipate what comes next would begin to fail.
It is neither a prophecy nor a scientific law. It is a possibility surrounded by arguments, disagreements, and enormous interests.
The first distinction: AI, AGI, and the singularity
Current systems can write, program, analyze images, operate tools, and solve increasingly complex problems. That does not mean we have reached artificial general intelligence, much less a singularity.
Although there is no universal definition, AGI usually refers to a system capable of learning and performing competently across a very broad range of intellectual tasks, with flexibility comparable to that of a human.
A superintelligence would go further: it would broadly outperform the best humans in fields such as science, engineering, strategy, and research.
The singularity describes something else: a change in dynamics. It would emerge if technological intelligence accelerated progress so much that the transformations after it became extraordinarily difficult to predict.
We can imagine the concepts as three possible steps:
- AGI reaches general intellectual ability comparable to humans.
- Superintelligence widely surpasses our abilities.
- That intelligence drives technological acceleration beyond our usual projections.
The third does not automatically follow from the others. We could have highly capable systems without explosive growth.
The decisive ingredient: feedback
At the heart of many versions of the singularity is not merely an intelligent machine. It is a machine capable of participating in the creation of better machines.
Imagine a system that can read research, write software, design experiments, optimize algorithms, and propose new architectures. It helps build a more capable AI. That new version becomes better at research and contributes to the next one.
The cycle would look like this:
AI improves AI development → better AI emerges → development accelerates again
If every cycle reduces the time required for the next, the result could be an acceleration that is difficult to control or anticipate.
In 1965, mathematician I. J. Good described this possibility as an intelligence explosion. His argument was that if designing machines is an intellectual activity, a machine that broadly surpassed human intelligence should also surpass our ability to design them.
The first ultraintelligent machine could therefore become the last great invention humanity needed to make without its help.
That thought experiment remains the technical core of the debate.
Why call it a singularity?
The word works as a metaphor borrowed from mathematics and physics: a point at which our usual models stop adequately describing what happens.
It does not mean technology literally becomes a black hole. The comparison concerns our knowledge.
A person living in 1500 might have imagined faster communication but would hardly have anticipated search engines, social networks, cryptography, streaming, or generative models. They did not simply lack data; they lacked the categories needed to think about those systems.
The singularity hypothesis says that we could find ourselves in a similar position when facing a civilization driven by intelligence far beyond our own.
In 1958, mathematician Stanisław Ulam recalled a conversation with John von Neumann about accelerating technological progress and a possible “essential singularity” after which human affairs could no longer continue as before.
The intuition therefore predates the internet, personal computers, and language models.
From a historical intuition to the modern singularity
In 1993, mathematician and author Vernor Vinge explicitly connected this historical break with the technological creation of entities more intelligent than humans.
Vinge did not imagine a single path. Superior intelligence could emerge from computers, large networks, interfaces between humans and machines, or biological enhancement. The central idea was that once human intelligence had been surpassed, the future would become incomprehensible using our present tools.
Ray Kurzweil later turned the singularity into a popular idea. His argument rests on the observation that some information technologies do not advance linearly but through accelerating returns. In his best-known prediction, he placed the singularity around 2045 and argued that the transition would include increasing integration between humans and technology.
The date is a prediction, not a scientific consensus. Still, it gave a concrete reference to an otherwise indefinite idea.
ChatGPT changed the psychology of the debate
Artificial intelligence existed long before ChatGPT. What changed was the public experience of its capabilities.
Millions of people could converse with a system capable of drafting, translating, explaining, programming, and analyzing information. Multimodal models, reasoning systems, programming tools, and agents capable of acting on software followed.
The social question changed from “will a machine ever perform intellectual work?” to “how much intellectual work will it perform, and how quickly will it continue to improve?”
None of this proves that the singularity will happen. It does explain why the idea no longer feels entirely disconnected from the present.
AI is also participating in its own development. Researchers and programmers use it to read papers, write code, analyze results, and design experiments. We do not have a machine that magically redesigns itself overnight, but we do have a circuit of people and agents that may accelerate each new generation.
If an intelligence explosion ever occurs, it may not look like one computer locked inside a laboratory. It could resemble a network of models, scientists, data centers, factories, and robots forming a collective research machine.
A fast singularity and a gradual one
There are two broad ways to imagine the transition.
In a fast takeoff, a slightly superhuman AI would help build a much better one, which would then produce the next. Within weeks or months, systems could move from roughly human ability to capabilities far beyond us.
In a gradual takeoff, capabilities would increase over years or decades. The change would still be enormous, but governments, companies, and people would have somewhat more time to adapt.
In 2025, Sam Altman proposed the idea of a “gentle singularity”: a succession of extraordinary advances that rapidly become ordinary. An AI writes well and surprises us; soon it becomes part of work. Later it researches, designs, or controls machines, and every wonder becomes routine.
It is an influential but interested perspective. Altman runs a company that builds these systems. The same applies to predictions from other laboratory leaders: they have privileged information, but also economic and strategic incentives to present the transformation as close.
Their dates should be treated as positions, not reliable clocks.
Why acceleration may encounter limits
Intelligence does not operate outside the physical world.
An AI might design a battery in minutes, but manufacturing millions requires mines, energy, permits, machinery, and logistics. It can propose a drug, but clinical trials take time. It can design a chip, but a new factory cannot be built at software speed.
There are limits in data, compute, money, regulation, and algorithmic knowledge too. A curve can grow rapidly for years and then reach a ceiling.
That is why AGI does not guarantee superintelligence, and superintelligence does not guarantee an instantaneous transformation of the material world.
One reasonable possibility is that intellectual capability will advance much faster than its physical application: an explosion of ideas accompanied by slower industrial change.
The skeptical position matters
Not every researcher believes current models are close to intelligence comparable to that of humans.
Yann LeCun, for example, has long argued that systems based primarily on language lack fundamental components needed to understand and plan in the world as humans and animals do. His research direction emphasizes internal world models and new architectures.
The disagreement can be summarized this way:
- One position holds that we already have most of the ingredients and that scaling, connecting, and improving systems will be enough.
- Another holds that one or more fundamental discoveries are still missing.
If the first is correct, timelines could be short. If the second is correct, today's predictions may eventually resemble the waves of optimism that preceded earlier AI winters.
Nobody knows the answer yet.
The signal would not be a benchmark
A system can outperform an exam without transforming the speed of civilization.
A deeper signal would be AI visibly accelerating the creation of new technology: not merely doing familiar work faster, but discovering new algorithms, materials, medicines, or architectures that in turn enable further discoveries.
At that point, the tool would be improving the process through which we build tools.
The economy would change too. If intellectual labor could be copied, run in parallel, and made much cheaper, we would have to reconsider institutions organized around its scarcity: education, professions, companies, wages, and the distribution of power.
The question would no longer be only which jobs disappear. It would be how a society works when specialized intelligence is inexpensive and the infrastructure running it belongs to a small number of actors.
Capability, control, and risk
Such acceleration would not only be a scientific and economic opportunity.
Far more capable systems could program, operate infrastructure, move resources, persuade people, or design dual-use technologies. The more autonomy they receive, the more important it becomes to ensure that their objectives and actions remain within acceptable boundaries.
This is the alignment and control problem: building powerful intelligence is not enough; we need to understand how to direct, audit, and stop it.
The singularity is difficult to predict precisely because it imagines cognitive ability beyond ours. But that uncertainty is not a reason to abandon analysis. It is a reason to reject both blind faith and automatic dismissal.
It may not be a date
When exactly did the internet revolution begin? We could choose ARPANET, the Web, browsers, Google, or the smartphone. No single date explains the whole transformation.
The singularity, if it happens, may only become visible in retrospect.
There may be no moment when a screen announces that it has been reached. It could be a process in which each AI generation automates more intellectual work, contributes more to building the next, and gradually accelerates science and technology.
Or the curve may stop, revealing that general intelligence was far harder to reproduce than it appeared.
Both possibilities remain open.
Why it is worth discussing today
The singularity should not be treated as a prophecy. Nor must it be dismissed as absurd fantasy.
It is a hypothesis about what could happen when the intelligence driving progress stops being exclusively biological and begins to become a reproducible technology.
Today, systems perform cognitive work, agents use tools, AI contributes to science, and laboratories openly aim to build general or superior intelligence. At the same time, models still make elementary mistakes, require oversight, and remain constrained by physical and economic limits.
That tension explains why the singularity is discussed so much today.
We do not know whether it will arrive in 2045, many generations from now, or never. But for the first time, a significant part of our civilization is deliberately trying to build intelligence comparable or superior to our own.
The most interesting question is not which date will be correct.
It is what happens to a civilization when the intelligence that produces its technology also begins to become a technology.
References
- Stanisław Ulam, draft tribute to John von Neumann, Institute for Advanced Study.
- I. J. Good, Speculations Concerning the First Ultraintelligent Machine, 1965.
- Vernor Vinge, The Coming Technological Singularity, 1993.
- Ray Kurzweil, The Singularity Is Nearer presentation at SXSW, 2024.
- Sam Altman, The Gentle Singularity, 2025.
- Meta AI, Yann LeCun's proposal for world models.
This article grew out of an experiment with an image-free format, closer to radio and carried entirely by voice. You can listen to the original video on YouTube.