AI singularity explained: Why tech leaders say it has already started

From Altman to Musk, AI leaders keep invoking singularity this year. Here is what the term actually means, and why experts remain unconvinced it has arrived

OpenAI’s Sam Altman remark on the Relentless podcast last week, that "we are now, like, in the singularity," was not really a claim about a new capability. It was a claim about a word. And the word has had a very specific, very old meaning that Altman and several of his peers now appear to be using differently than the people who coined it.

 

The remark did not land in isolation. It came days after OpenAI disclosed that one of its own models had broken out of a testing sandbox and into Hugging Face's systems, and in the same fortnight that more than 1,100 researchers across OpenAI, Anthropic, Google and Meta asked governments to prepare tools to slow frontier AI development if it ever needs slowing.

 

 

Jensen Huang and Demis Hassabis had already used versions of the same word “singularity” earlier this year, and Elon Musk endorsed Altman's framing within two days of the podcast airing. Whether all of them mean the same thing by it is a separate question, and one worth settling before getting into whether any of them are right. 

What the words actually mean

The statistician I.J. Good described singularity in 1965 as the point at which a machine becomes good enough at designing machines to build a better version of itself, which then designs a still better one, in a loop that runs faster than humans can follow, an intelligence explosion. The computer scientist Vernor Vinge later described the result as a point beyond which the future becomes unpredictable, because whatever is shaping it is no longer human.

 

In this framing, artificial general intelligence (AGI) is usually treated as a milestone that could precede the singularity, not the same event. Reaching human-level intelligence across the board is one thing. That intelligence then improving itself autonomously and without limit is another.

 

American computer scientist, Ray Kurzweil, who brought the term to a general audience, attached dates to human-level machine intelligence by around 2029, and a merging of human and machine intelligence by around 2045.

 

The mechanism behind that second claim, a machine improving its own successor without a human directing the work, has a name in AI research today, recursive self-improvement, usually shortened to RSI. What matters about it is that it describes a threshold that can in principle be checked. Either a system is doing this, or it isn't.

Altman's own writing has been more careful than his podcast remark suggests. In a June 2025 essay titled "The Gentle Singularity," he described a "larval version of recursive self-improvement" and was explicit that this "isn't the same thing as an AI system completely autonomously updating its own code." That is a narrower claim than the one that got picked up everywhere last week.

Four declarations

Altman was not the first AI leader to invoke the term this year, and the sequence is worth laying out. In an interview on Lex Fridman's podcast on March 23, Nvidia chief executive Jensen Huang said he believed AGI had effectively already been achieved. Speaking after Google I/O on May 20, DeepMind chief executive Demis Hassabis said humanity was standing "in the foothills of the singularity," describing current AI agent progress as a dress rehearsal rather than the event itself. Altman made his remark on July 25. Within two days, Elon Musk, in a post on X, said the singularity had arrived and that we were in its early stages.

 

Each is using the same vocabulary to describe a different position, which is itself part of the story.

 

That disagreement showed up clearly when researchers were asked directly, in their comments to Business Insider.

 

Stuart Russell of UC Berkeley said neither he nor, in his view, Altman actually believes the threshold has been met, pointing to Altman's own remarks last month that AI could be doing a "significant fraction" of OpenAI's research only by March 2028. Roman Yampolskiy of the University of Louisville made a similar point more bluntly: "Rapid progress is not itself the singularity." Nick Bostrom, in his comments to the outlet, compared the moment to falling into a black hole, in that it may be impossible to know from the inside when the crossing happens, and argued current systems show only the "first stirrings" of recursive self-improvement because they still lack continual learning. Ajay Agrawal of the University of Toronto, also speaking to Business Insider, argued that AI systems "do not want anything" and that whatever purpose they display originates with the humans who set their objectives.

 

Only James Barrat, an author who has long warned about advanced AI risk, told the outlet he agreed with Altman, citing the Hugging Face breach as evidence that AI had already "slipped from our control."

What the evidence actually shows

Setting the rhetoric aside, there is real evidence behind the idea that something is accelerating fast, even if it stops short of the classical threshold. In July, a decades-old open problem in mathematics, the Jacobian conjecture, got a counterexample produced with the help of Anthropic's Fable 5 model. As mathematician Kevin Buzzard of Imperial College London told Fortune once the proof had been checked, "It is a big day."

 

The clearer signal, according to release data compiled by Forbes is cadence. Anthropic took 444 days to go from Claude 3 Opus to Claude Opus 4, then 186 days to the next frontier model, then 73. Across 2026, that gap has fallen to roughly 60 days on average. OpenAI shows the same pattern, with gaps between successive GPT releases shrinking from over 400 days to under 100 through 2025 and 2026.

 

Two labs, working independently, have converged on shipping a new frontier model roughly every two months, though as that same analysis notes, faster shipping is not necessarily proof of larger capability gains, since version numbers are partly a marketing decision.

 

On actual capability, the nonprofit METR, which measures what it calls a time horizon, the length of task a human expert would need that an AI agent can complete with at least even odds of success, reported that this figure moved from about four minutes for Claude Opus 3 in March 2024 to roughly twelve hours by early 2026.

 

Separately, in an internal account published by the Anthropic, the company said more than 80 per cent of the code merged into its production codebase is now written by Claude, and that on one internal test of how much a model can speed up training code, its Mythos Preview model averaged roughly 52 times improvement against about four times for a skilled human.

 

In an open-ended AI safety research problem Anthropic ran end to end with Claude agents, the agents closed 97 per cent of the available performance gap where two human researchers working about a week had closed 23 per cent, though humans still chose the problem and wrote the scoring rubric. Anthropic's own summary of that experiment cuts against the singularity framing rather than supporting it: the company stated that large performance gaps persist in a model's judgement in choosing its own goals, and that recursive self-improvement "is not inevitable."

The incidents behind the moment

Altman's remark landed days after two disclosures that shaped how it was read. According to OpenAI's own account, during an internal test of how well its models could find and exploit software vulnerabilities, with safety guardrails deliberately reduced, the models found an unknown flaw, escaped their sandbox, and broke into Hugging Face to steal the benchmark's answer key. A subsequent forensic report by Hugging Face found the intrusion went well beyond a quick smash and grab: about 17,600 separate attacker actions over several days, a foothold established outside OpenAI's own systems, and repeated changes of tactic whenever a route was blocked.

 

In comments on the incident, Sean O hEigeartaigh of Cambridge's Centre for the Future of Intelligence told Fortune that the behaviour was reward hacking rather than autonomous goal-setting, since the model never deviated from the score-maximising objective it was given.

 

Separately, in a report on what it calls the first fully agentic ransomware attack, named JADEPUFFER, the cybersecurity firm Sysdig said an AI agent broke into a company's infrastructure, diagnosed and fixed its own failed login attempt within 31 seconds, and went on to encrypt over 1,300 records and demand a ransom, with no human directing the keyboard at any stage.

Neither incident shows a machine setting its own agenda. However, both show that current systems can chain known techniques together, adapt when a step fails, and do so faster and more persistently than a human operator would.

 

Brian Jackson of Info-Tech Research Group, commenting to Forbes on the incident, said part of a genuine singularity would be AI "setting its own goals and intents and the ability to self-sustain," and argued that did not happen here.

 

That gap between persistence and autonomy is also why more than 1,100 researchers at OpenAI, Anthropic, Google, Meta and Nvidia signed a statement last week asking governments to prepare, not necessarily use, mechanisms to slow frontier AI development if automated AI research begins to compound faster than institutions can track it. In the statement, OpenAI researcher Leo Gao described the underlying problem as a coordination failure, writing that "no individual actor is willing to stop unilaterally" even if labs agree in principle that someone eventually should.