Insights

What's Actually Behind the AI Singularity

In physics, singularity means: equations break down here. Whether and when AI reaches that point — Kurzweil, Hinton, LeCun and Marcus are decades apart. A search for what can actually be measured.

10 September 2026 · 13 min read AIFutures Auf Deutsch lesen Share on LinkedInShare on X

In physics, a singularity is a point where an equation breaks down, where a function that's been perfectly smooth suddenly stops returning a meaningful value, just "divide by zero." The center of a black hole is a point like that: infinite density, where any physics that worked up to that point simply stops applying. That's exactly the image computer scientist and science-fiction author Vernor Vinge borrowed in 1993 when he coined the term for a possible technological future: the moment human intelligence gets overtaken by a self-improving machine intelligence is, for Vinge, an event horizon beyond which prediction stops working, "similar in some sense to the knotted space-time at the center of a black hole."

That's a bold claim. And a quick look at who believes it today, and who doesn't, makes one thing clear fast: there's no agreement, not even close. Ray Kurzweil has dated the event to 2045 for 30 years, unchanged. Geoffrey Hinton, one of the three "godfathers of deep learning" and a Nobel laureate, considered superintelligence 20 to 50 years away as recently as two years ago; today he considers "maybe even five years from today" plausible. Yann LeCun, Meta's chief AI scientist, calls the exact technology OpenAI and Google are building a detour, not a path to the destination, and just backed the counter-bet with a billion dollars. Gary Marcus thinks neither 2027 nor 2050 is realistic, more like 2100, if ever. Four compass needles, four different directions. That's the actual point of this piece: not who's right, but why people this smart can disagree this much.

What "singularity" originally meant

The term doesn't come from Kurzweil. In 1965, mathematician I. J. Good described, in "Speculations Concerning the First Ultraintelligent Machine," a first ultraintelligent machine that, once built, could design even better machines itself: "there would then unquestionably be an 'intelligence explosion.'" Picture it as a chain reaction, not one improvement, but an improvement that triggers the next one, which triggers the one after that, each round faster than the last. In 1993, Vinge poured that idea into the term we use today, with the event-horizon image mentioned above. He dated the event to somewhere between 2005 and 2030.

Worth noting for context: singularity, in its original sense, is a more specific, harder claim than AGI (human-level capability across most tasks) or superintelligence (intelligence far surpassing even the brightest humans). Superintelligence is a possible outcome of the singularity, not the singularity itself, the actual chain-reaction event after which prediction fundamentally stops working. Even Kurzweil's own model reflects this distinction: he predicts AGI for 2029, but the chain reaction itself not until 2045, 16 years later. Even the industry's most optimistic forecaster leaves an entire generation between "the machine is as smart as we are" and "the machine explodes past us intellectually."

The mechanism behind it: snowball or sprinter?

Here's the real, rarely asked question: why would intelligence improve itself explosively at all, instead of just getting better step by step? Two images help make sense of the two camps.

The first is a snowball rolling down a hill. Small at first, but every rotation picks up more snow, and more mass means it rolls faster, which picks up even more snow. The feedback loop reinforces itself. Applied to AI: an AI smarter than its human developers could do AI research better than they can, building the next, even smarter AI faster, which then builds the one after that faster still. That's the core of what AI safety research has called the "foom" scenario ever since a famous online debate between economist Robin Hanson and AI theorist Eliezer Yudkowsky in 2008, named for the sound of a rocket launching: a fast, local, barely stoppable breakout.

The second image is a sprinter closing in on their personal best. Early on, shaving time off is easy, every training session shows real gains. But the closer you get to the physiological ceiling, the smaller the gains get, and the more effort the same improvement costs. Hanson's counter-position in that very debate, and LeCun's and Marcus's position today, follows this image: progress gets more expensive, not cheaper, the further you go, held back by real constraints like training data, energy, or an architecture that simply doesn't learn what intelligence actually consists of.

Almost 20 years after that debate, it's still unresolved. That's the actual bet underneath the entire singularity debate, more important than any single calendar year.

Altman's reframing: the "gentle" singularity

In June 2025, Sam Altman published an essay titled "The Gentle Singularity" that deliberately recharges the term. His thesis: "the singularity happens bit by bit, and the merge happens slowly. We are climbing the long arc of exponential technological progress; it always looks vertical looking forward and flat going backwards, but it's one smooth curve." In the same piece, he nonetheless dates concrete capability jumps: "2025 has seen the arrival of agents that can do real cognitive work [...] 2026 will likely see the arrival of systems that can figure out novel insights. 2027 may see the arrival of robots that can do tasks in the real world."

Vinge's singularity is a light switch: on, off, and after that it's dark, meaning unpredictable. Altman's "gentle" version is a dimmer: slowly brightening, always visible, never a break. Except his dimmer uses suspiciously precise years for something that supposedly has no break at all. Both texts use the same word for two nearly opposite claims, a distinction rarely carried into the public debate, or into Altman's legal fight with Musk, where exactly this question, how fast and how controlled the journey really is, sits at the core.

What can actually be measured

Part of this debate is pure opinion. Another part is measurement. The research organization METR has tracked, since 2019, how long a task is allowed to take, measured against the time a human expert would need, for an AI model to still solve it at a 50% success rate. GPT-2 landed at roughly two seconds. Claude 3.7 Sonnet at 50 minutes. By early 2026, Opus 4.6 reached roughly twelve hours. Between 2019 and 2025, that figure doubled on average every seven months; between 2024 and 2025, the pace accelerated to a doubling every four months. If that trend holds, month-long tasks become reachable starting in 2027, the same snowball from the section above, this time in measurements instead of metaphors.

Before trusting that curve blindly, it's worth looking at the most famous exponential curve in tech history. In 1965, Gordon Moore predicted the number of transistors on a chip would double every year; in 1975, he revised that to every two years. That curve genuinely held, for more than 40 years, until the chip industry officially dropped the formula from its own roadmaps around 2016, as physical limits (heat, leakage current, atomic spacing) caught up with it. Moore's Law is the rare case of an exponential curve that actually held, long enough to shape an entire industry, but not forever. That same double lesson, take it seriously and stay skeptical anyway, applies just as much to METR's curve.

The counter-position: real limits, not excuses

This is exactly where the serious counter-case starts, and it deserves more weight than most headlines give it. LeCun's argument is architectural, not rhetorical: language models learn from text, a thin slice of how intelligence actually works, without embodied experience, without a cause-and-effect understanding of the physical world. In May 2026, he said on Bloomberg, verbatim: "Large language models are not the path to real intelligence. They're a detour." In March 2026, his new venture AMI Labs closed a $1.03 billion seed round, with the stated goal of building world models instead of language models, the largest institutional bet yet that pure scaling of language models won't get to AGI: the same sprinter from above, who needs a different training method, not faster running shoes.

Gary Marcus isn't making a simple denial. His often-misread 2022 thesis, "Deep Learning is Hitting a Wall," was never the claim that AI can't do anything, only that scaling alone isn't enough. In 2026, he reaffirms: "Anyone who thinks AGI is impossible: wrong. Anyone who thinks AGI is imminent: just as wrong." His own timeframe isn't 2027, more like beyond 2050, possibly not until 2100. Worth noting: a Manifold prediction market asking explicitly whether the "deep learning hit a wall" reading would become the accepted view by year-end stood at just 4% YES in August 2026; the market itself mostly doesn't buy Marcus's framing.

Perhaps the weightiest case comes from Ilya Sutskever. Few people did more to prove out the scaling paradigm behind GPT-3 and GPT-4 than OpenAI's former chief scientist. With his new venture, Safe Superintelligence, he now argues that the age of pure scaling is over, and is betting on new research directions instead. Notably contradictory: in July 2026, SSI signed a partnership with Nvidia that increases its available compute by an order of magnitude, no retreat from fuel, then, but a rebuild of the engine that burns it.

The limits themselves aren't pure opinion either. Per Epoch AI's analysis, the stock of high-quality, publicly available text training data is finite, roughly 300 trillion tokens, and models could run that tank dry between 2026 and 2032, a window that has just begun. Synthetic and multimodal data will likely soften that constraint, a second tank of sorts. Power and chip-manufacturing capacity count as the harder bottlenecks in the same analysis, but those have so far kept shifting under ever-larger capital investment, not a law of nature, then, but a question of capital and time.

The graveyard of postponed futures

Anyone who distrusts the optimists doesn't have to rely on gut feeling. There's a well-documented precedent, and AI research itself has lived through it twice already.

In 1965, future Nobel laureate Herbert Simon wrote: "Machines will be capable, within twenty years, of doing any work a man can do." Twenty years later, in 1985, machines weren't remotely close. The overoptimism wave of the 1950s and 60s ended instead in the first "AI winter" of 1974: in Britain, the Lighthill Report painted a bleak picture of AI research and cut funding there, and in the United States, the research agency DARPA pulled its own funding for general AI research around the same time, three million dollars in annual grants at Carnegie Mellon University alone, cut overnight. A second AI winter followed from 1987 to 1993, as the market for specialized AI hardware collapsed once ordinary computers suddenly matched its performance for a fraction of the price.

Nuclear fusion tells the same story from a different field: since the 1970s, it's been "30 years away," fairly consistently, decade after decade, a running joke even the head of the UK's Atomic Energy Authority openly admits to. The pattern is always the same: a real, demonstrable improvement curve gets extended linearly, or exponentially, into the future, and the rest of the curve turns out considerably flatter than expected, sometimes for decades.

That doesn't mean it has to go the same way this time. Moore's Law, in the section above, supplies the counter-example: some exponential curves genuinely hold, long enough to change the world. It just means "this time it's different" has been the default claim of every generation of tech optimists, in this very field, with similar arguments, twice in one lifetime.

What the forecast landscape itself reveals

One detail deserves attention on its own: how unstable the forecasts are over time. A survey of AI researchers found a 2022 median of 2060 for "high-level machine intelligence"; by 2023, that same median had slipped to 2047, a 13-year shift within a single year. A survey of professional forecasters and superforecasters put the odds in February 2026 at 25% for AGI by 2029 and 50% by 2033, with both figures already pushed back by two years from the previous year's estimates. A third survey, run in May 2026 (LEAP), asked more concretely about a specific capability threshold on software tasks: experts landed on a median of 2030, superforecasters on 2028, earlier than the experts themselves, while the general public came in at 2037.

It's like watching a compass that isn't broken, just honestly reporting that the magnetic field itself is drifting. Three groups, three systematically different answers to the same question, none of them stable over twelve months. That's the single most telling observation in the whole debate: naming a year today captures a snapshot, not a certainty.

What this means for founders and business owners

From my work with 30+ founders, I draw a very concrete conclusion: building your house on a fortune teller's date, whichever one, is the actual mistake, not picking the wrong year. When Nobel laureates, chief scientists at major labs, and professional forecasters disagree by decades on exactly this question, and when this exact field has already overestimated its own future twice, "I think it happens in 2029" is not a foundation for a three-year roadmap.

What can be planned, though, is the METR trend itself, applied to your own domain instead of abstract benchmarks: what task length can an agent reliably handle in your specific use case today, and how has that figure moved over the last six months? That's the same logic as the narrow, measurable scope from article one in this series, just repurposed as an early-warning system for your own capacity planning: not whether or when the chain reaction arrives, but how fast the capability ceiling relevant to your business is actually moving right now. You only ever see the event horizon itself once you're already past it. Building your business on ground you can actually survey today is the one strategy that works regardless of whether that horizon is three years out, or seventy.

Sources:

Historical concepts

  • Wikipedia: Technological singularity
  • I. J. Good: "Speculations Concerning the First Ultraintelligent Machine" (1965)
  • Vernor Vinge: "The Coming Technological Singularity" (1993)

The mechanism debate

Forecasts

Counter-position

Historical precedents

Empirical data & prediction markets

Direct follow-on

More Insights
Albert Schaper
About the author

AI expert, entrepreneur, and founder with a finance background and a focus on execution. I build companies, invest in ventures, and advise teams on putting AI to work. LinkedIn

← All Insights