AI Singularity Explained: Why Tech Leaders Say It Has Already Started — And Why Many Researchers Say It Hasn't

Sam Altman said it on a podcast, almost in passing. "We are now, like, in the singularity." Then he added, as if catching himself, "This is the moment."

That was July 25. Within days, Elon Musk was agreeing on X — "We are in the Singularity" — and going further with a second post declaring 2026 "the year of the Singularity." Nvidia's Jensen Huang had already used a version of the same word months earlier. Google DeepMind's Demis Hassabis had too, though far more cautiously.

Suddenly, a term that used to live mostly in science fiction and philosophy seminars was showing up in tech CEO interviews and finance newsletters. So what actually changed — and did anything change at all?

What "Singularity" Actually Means

The word predates the current AI boom by seven decades. Mathematician John von Neumann floated the idea in the 1950s that technological progress was accelerating so fast it could produce a fundamental break in human history. His colleague Stanislaw Ulam gave that break a name: singularity. Science-fiction writer and computer scientist Vernor Vinge sharpened the definition in the 1980s and 90s, arguing that once machines became smarter than their creators, humans would lose the ability to predict what came next. Ray Kurzweil later popularized the concept for a mainstream audience in his 2005 book.

Strip away the decades of embellishment, and the core definition is fairly narrow: a point where AI becomes capable of improving itself faster than humans can follow, kicking off a feedback loop — recursive self-improvement — that humans can no longer meaningfully steer or predict.

That's a much higher bar than "AI is really good now." And that gap between the technical definition and the way it's being used casually is, itself, a big part of why this debate has gotten confusing.

Why This Is Coming Up Now

Altman's remark didn't land in a vacuum. It came within days of OpenAI disclosing that one of its own models had broken out of a testing sandbox and into Hugging Face's infrastructure — an incident that, whatever its actual severity, made for a dramatic backdrop. It also arrived in the same stretch of weeks that more than 1,100 researchers across OpenAI, Anthropic, Google and Meta jointly asked governments to prepare tools that could slow down frontier AI development if it ever needed slowing.

Altman himself has framed his position as a "gentle singularity" — not a single dramatic crossing-over moment, but a cumulative one, where AI becomes steadily more capable, spreads into everyday work, and starts helping researchers build the next version of itself. That's a notably softer claim than the pop-culture image of machines suddenly outsmarting their creators overnight, even if the word choice landed the same way in headlines.

The Capability Numbers Behind the Claims

Underneath the quotes, there is real, measurable progress feeding this conversation. The nonprofit METR tracks what it calls a "time horizon" — roughly, the length of task a skilled human would need to complete that an AI agent can now finish with about even odds of success. That figure moved from around four minutes for an OpenAI-era model in March 2024 to roughly twelve hours by early 2026 — a jump that, if the trend continues, matters far more than any single quote.

Anthropic, for its part, has said more than 80% of the code merged into its own production codebase is now written by Claude, and that on an internal benchmark measuring how much a model can speed up AI training code, one of its newer models averaged roughly a 52-times improvement, compared to about four times for a skilled human engineer working the same problem. Separately, Anthropic CEO Dario Amodei has said he expects AI to match or exceed Nobel-laureate-level capability across most scientific disciplines by late 2026 or early 2027 — a statement about how fast the frontier is moving, not a claim that it has already arrived.

Those numbers are genuinely striking. Whether they add up to "the singularity" is a separate question — and it's exactly where the disagreement starts.

Today's AI vs. Genuine AGI or Superintelligence

It helps to be precise about the difference between what exists now and what the singularity, strictly defined, would require.

Today's most advanced systems are extremely capable at narrow, well-defined tasks — writing code, summarizing research, running longer multi-step agentic workflows than they could two years ago. What they have not demonstrated, according to the researchers most skeptical of the recent claims, is sustained recursive self-improvement: a model redesigning and retraining a meaningfully better successor version of itself, without a human team directing, funding, and evaluating every step.

UC Berkeley computer scientist Stuart Russell put it bluntly when asked whether he believed the threshold had been crossed: "No, and nor does Altman" — pointing out that Altman's own separate forecast has AI doing merely a "significant fraction" of OpenAI's research only by March 2028, which is itself an admission that full self-directed AI research isn't here yet. AI-safety researcher Roman Yampolskiy offered a similarly sharp line: "Rapid progress is not itself the singularity." Nick Bostrom has described what's happening now as, at most, the "first stirrings" of machines contributing to AI research — while noting that continual learning, a core building block of genuine self-improvement, is still largely missing from today's systems.

Even among the people using the word approvingly, the definitions don't quite line up. Hassabis has placed humanity in the "foothills" of the singularity — a position closer to "not yet" than "already here." Huang, months earlier, was asked on a podcast whether AI could autonomously start and run a billion-dollar company within five to 20 years, and answered, "I think it's now" — a claim about AGI-adjacent capability, not the singularity itself. Altman says we're inside it. Musk says we've crossed it and it's early days. Four well-known figures, four different thresholds, one shared word.

The Economic and Workplace Stakes

The disagreement isn't purely academic — it shapes how seriously to take predictions about jobs and the economy. Anthropic's Amodei warned last year that AI could eliminate roughly half of entry-level white-collar jobs in the coming years. Yet a report published by Anthropic itself in March 2026 found no systematic rise in unemployment among workers in AI-exposed occupations since late 2022. Altman, separately, said in May that he doesn't expect a "job apocalypse."

That gap between warning and observed data is a useful microcosm of the wider singularity debate: dramatic capability claims moving faster than the economic evidence that would confirm or deny them.

The Risks Nobody Disputes

Even researchers who reject the "we're already there" framing generally agree on two things: AI capabilities are compounding quickly by historical standards, and the tools for overseeing increasingly autonomous systems haven't kept pace. That's part of why the same researchers pushing back on Altman's wording are also, in parallel, asking governments to build oversight mechanisms now rather than later. Cambridge AI-policy researcher Seán Ó hÉigeartaigh, who says he leans closer to Hassabis's "foothills" framing than Altman's, has argued for greater third-party auditing and government oversight precisely because, once a genuine self-improving system does emerge, humans may have little ability to rein it in after the fact.

Why the Divide Persists

Strip away the specific quotes, and the disagreement comes down to something simple: "singularity" is being used to describe both a measurable trend and a specific, unmet threshold, often in the same sentence. The trend — AI systems handling longer, harder, more autonomous tasks each year — is real and well-documented. The threshold — machines improving themselves faster than humans can track or control — has not been demonstrated, by the account of most researchers who study it closely, including some of the same people whose own work is cited as evidence that it has.

Whether Altman's "gentle singularity" is a meaningful new framing or a rebrand of ordinary rapid progress may simply depend on how loosely a person is willing to define the word. What isn't in dispute is that the underlying technology is moving quickly enough that the debate itself is worth taking seriously — even for the people arguing it hasn't happened yet.