Sam Altman says: “We're already in the singularity”. What it really means and why the most important question is a different one

“We are in the singularity. This is the moment.” With that sentence — delivered almost casually, with no triumphant announcement and no elaborate press conference — OpenAI CEO Sam Altman described the present as something that ten years ago still looked like academic science fiction: a distant hypothesis debated by philosophers and futurists at the margins of mainstream science. The statement immediately produced two opposite reactions in the tech community: those who read it as an announcement that machines have already surpassed human intelligence, and those who dismissed it as marketing hyperbole from an entrepreneur with a personal stake in AI hype. Both reactions miss the point. The question is not whether Altman is right. The question is: if the singularity really were a process already under way, would we be able to recognise it while it happens? That is the question worth exploring, because the answer says something important not only about AI, but about how humans recognise historical change while living through it.

First it matters to understand what Altman means by the word, because “singularity” carries a precise conceptual baggage that he is deliberately using differently from the classic definition. The term “technological singularity” was popularised by mathematician and writer Vernor Vinge in his 1993 essay and taken mainstream by Ray Kurzweil in “The Singularity Is Near” (2005). In its classic formulation it is the point at which artificial intelligence surpasses human intelligence across all relevant dimensions, after which technological progress accelerates so fast it becomes unpredictable and incomprehensible to humans: a discontinuity, like a mathematical singularity, beyond which prior models stop working. That definition implies a precise, recognisable moment — the day a machine beats a human on test X, benchmark Y, task Z. A sharp boundary.

Altman uses the term in a meaningfully different sense, the one he had already called a “gentle singularity” in earlier interviews. Not a sudden explosion, no cinematic moment in which a machine wakes up and decides to take control, no calendar date we will look back on as the day everything changed. Instead a curve: a gradual process of acceleration already in motion, which we have already entered without any single recognisable turning point. Capabilities increase progressively, enter daily life, and what looked impossible yesterday becomes first surprising, then normal, then obvious. No cinematic moment, no machine suddenly awakening — only a continuous acceleration that, if you pay enough attention, you can already recognise by looking back two or three years.

At the heart of Altman's statement is a precise technical observation, not just a philosophical metaphor: AI today writes code, accelerates scientific research and helps the very engineers who are building even more advanced models. Progress therefore begins to feed further progress. That is the key mechanism — not the absolute intelligence of any single system, but the feedback loop between AI and AI development itself. Until a few years ago the process was linear: human researchers designed architectures, wrote code, ran experiments, analysed results, published papers, other researchers read and improved. A slow cycle, bounded by human cognitive capacity and available time.

Today the cycle has changed structurally on three fronts. First, AI models accelerate AI research: GPT-5.6, Claude Fable 5 and Kimi K3 are used daily by the teams building the next generation — analysing papers, suggesting hypotheses, writing experimental code, spotting bugs, explaining counterintuitive results. Researchers who use advanced AI to do AI research are measurably more productive than those who don't. Second, AI models write code for themselves: a growing share of the code implementing the newest AI systems was written or substantially assisted by AI. We are not yet at the point where a model autonomously reprograms itself, but the distance to that scenario is shrinking. Third, AI models accelerate the collection and analysis of training data — one of the main bottlenecks in AI development — by generating high-quality synthetic data, finding and fixing errors in existing datasets, and classifying unstructured data.

When the system that needs to improve actively contributes to its own improvement, the rate of progress accelerates. Not indefinitely — physical and computational limits exist — but enough to produce what Altman calls the signal of the singularity: progress feeding further progress.

The most challenging idea in Altman's statement is that the great historical transition may not have a precise date. Not a moment we will remember as “the day AI changed everything”, but a gradual process we could pass through without noticing, recognising it only afterwards, when the previous world suddenly feels distant. Think of the internet: when exactly did it change the world? There is no date. Not 1969 with ARPANET, not 1991 with the World Wide Web, not 1995 with Netscape. Change happened gradually until, in retrospect, the world before the internet looked like another era. How many people in 1997 would have said “yes, we are living through an epochal economic transformation right now”? Very few: most saw slow websites, attachments that took minutes to download and dial-up connections tying up the phone line. The big picture was invisible from the inside.

The same holds for the printing press. Gutenberg invented movable type in 1440, the Protestant Reformation began in 1517, the Scientific Revolution took hold over the 16th and 17th centuries. The causal connection is obvious in retrospect — print democratised access to information and made radical cultural and political transformations possible — but nobody in 1440 could have predicted the outcome. Someone standing inside the printing singularity in 1460, as the first Gutenberg Bibles began circulating, did not recognise it as such: they saw a craftsman producing books faster. The epochal transformation was invisible from the inside.

Are we in an analogous situation? Altman suggests we are, and his argument is hard to dismiss lightly. In July 2026 AI writes a significant share of the code produced globally, accelerates scientific research in biology, chemistry, physics and mathematics, handles millions of customer-service interactions a day, helps professionals complete complex tasks in a fraction of the previous time, and actively contributes to building the next generation of models. Seen from the inside, as we see it every day, it feels normal: the natural result of years of technological progress. Seen from the outside — from the perspective of someone living in 2016 who could watch 2026 — it would look extraordinary. Almost impossible. Almost like a singularity.

The statement does need some qualifications, or it is easily misread. First: this does not mean AGI has arrived. Current systems, impressive as they are, still make elementary mistakes on tasks any child would handle, fail on long and complex work requiring extended planning, sustained attention and adaptive error correction, and show fundamental gaps in deep causal understanding of the physical world. Artificial general intelligence — a system matching or exceeding human cognitive capability across all relevant dimensions — is not reality in July 2026: researchers have no consensus on when it will arrive or how to recognise it, and the definitions of general intelligence themselves remain deeply contested. Altman did not announce AGI. He described an acceleration process — the singularity as a curve, not as a threshold crossed.

Second: benchmarks are not reality. Models that beat human scores on standardised tests are not equivalent to humans in real life. A model above the 99th percentile on a maths test can fail completely at practical tasks a thirteen-year-old would do without thinking: generalisation remains the Achilles heel of current systems. Third: acceleration is not infinite. The singularity as a mathematical concept implies acceleration tending to infinity, but reality is more prosaic — physical, computational, energy and cognitive limits put real ceilings on progress. The curve can be steep, but it is not vertical.

Which brings us to the most important question of all: if the singularity really were already under way, would we be able to recognise it while it happens? The honest answer is probably not, and there are good cognitive and structural reasons why. The first is progressive normalisation: humans adapt fast to the new normal. Every time a new capability ships — the first time ChatGPT wrote a university-level essay, the first time a model generated photorealistic images, the first time one solved olympiad maths problems — the initial reaction is astonishment. Six months later it is normal, a year later it is taken for granted, two years later it seems obvious machines can do it. That normalisation makes it structurally hard to perceive accumulation: each new normal is far more advanced than the last, but the cumulative trajectory stays invisible moment to moment.

The second reason is comparison with the immediate past. We judge change against the recent past, not the distant one. If a model in July 2026 is 20% more capable than January 2026's, the subjective perception is “incremental improvement”; compare it with what was available in July 2023 and the difference would look impossible. The singularity, if it is happening, hides precisely in that bias: measuring change against the present instead of a long enough horizon. The third reason is the expert-silo problem: researchers see the limitations of current models every day, know better than anyone where they fail and how far AGI is in the technical sense, and that makes them naturally sceptical of claims like Altman's. Non-experts — the doctor using AI to support a diagnosis, the engineer who has it write half the code, the small-business owner automating customer support — feel the jump far more directly, because they measure it against their own work rather than benchmarks. Neither perspective is complete on its own.

What does this mean in practice for anyone running a company or building products? Three concrete things. First, stop waiting for the event. If the singularity is a curve and not a threshold, no announcement will tell you when the right time to adopt AI has come — the right time is always “now, with the latest state of the art”. Second, measure over long horizons. Every six months, re-evaluate the use cases you dropped because “AI can't do it”: a significant share of those noes turns into yes without anyone notifying you. Third, build for model substitutability. If capabilities grow and prices collapse every few months, the winning architecture is the one where the model is an interchangeable component, not the immovable core of the system.

In the end, Altman's statement is less a technical announcement and more an invitation to change our unit of measurement. Whether he is right or wrong about the word “singularity” matters relatively little: what matters is that the pace of change has outrun our ability to perceive it in real time. And that is the most concrete risk for people at work — not AI taking control, but realising too late that the world around you has already changed the rules. If there is one thing worth taking away, it is this: don't ask whether we are in the singularity. Ask what you would do differently today if we were.