Musk and Altman say AI has hit the singularity — but how would we know?

Full text

Artificial intelligence (AI) may have already reached the "singularity" ‪—‬ the long-theorized threshold beyond which humans cannot predict the technology's advancement, prominent AI executives such as Elon Musk and OpenAI founder Sam Altman have said.

In a post on social media platform X , Musk pointed to a number of recent incidents of AI systems exceeding their previously assumed limits, including hacking external systems and completing previously unsolved math problems .

First conceptualized by mathematician and Manhattan Project scientist John von Neumann in the 1950s and popularized by science fiction, "the singularity" refers to an inflection point beyond which the evolution of technology becomes impossible for humanity to predict or control.

While definitions vary, the term is most often applied to AI, with the arrival of artificial general intelligence (AGI) viewed as its primary catalyst. The emergence of AGI — a future AI system that can exhibit human-level cognitive function and reasoning across any discipline rather than a specifically trained subset — represents a significant milestone for the technology.

It is the point at which an AGI system could recursively improve its own capabilities, which some say will trigger artificial superintelligence (ASI) as it moves along an exponential curve and quickly exceeds the intelligence of its creators.

Elon Musk has pointed to a string of recent AI achievements in math and computer science as evidence for reaching the singularity. But experts aren't convinced. (Image credit: Andrew Harnik via Getty Images) According to a 2025 study that analyzed over 8,000 predictions from AI experts, entrepreneurs and scientists, some believe there is a roughly 50% probability that human-level AGI will be reached within a few decades . Some figures ‪—‬ including Google DeepMind co-founder and chair Demis Hassabis ‪—‬ suggest we're in the early stages of the singularity already.

In his 1993 essay, " The Coming Technological Singularity: How to Survive in the Post-Human Era ," sci-fi author and mathematician Vernor Vinge — who was also one of the first to explore the concept of cyberspace — examined the question of how the singularity might manifest and the symptoms by which society might recognize its imminent arrival.

"Since it involves an intellectual runaway, it will probably occur faster than any technical revolution seen so far," he wrote. "The precipitating event will likely be unexpected —- perhaps even to the researchers involved."

Breaking boundaries

Unexpected events have been rife in the AI community in the past several months. Anthropic representatives revealed on July 30 that the company's AI model Claude broke out of its locked-down testing environment during a security evaluation and hacked multiple external organizations. The report followed a similar incident in which an unreleased OpenAI model broke containment and hacked into AI training repository Hugging Face.

In these examples, experts said the containment breach and subsequent hacks happened because the models were trying to fulfill their prompts as efficiently as possible. Earlier this year, Anthropic's Project Glasswing also demonstrated an ability to discover and map thousands of previously undetected zero-day cybersecurity vulnerabilities, while other AI models have disproved or solved a range of previously incomplete math problems .

However, Jon Crowcroft , a professor of communications systems at the University of Cambridge and a researcher at The Alan Turing Institute, says that this is less a sign of a technological tipping point and more a problem of proper configuration.

"To be honest, that was incompetence on both sides — they claimed the AI was being trained in the ExploitGym , but that just means it wasn't properly sandboxed," he told Live Science in an email. "Sandboxing is something we do all the time to stop this sort of exfiltration and infiltration." For example, Crowcroft said he and colleagues designed a system for the U.K.'s National Health Service to safely work on confidential data behind double firewalls, and for years, the Financial Conduct Authority (the U.K.'s financial services regulator) has had a system for running algorithmic traders in a safe sandbox.

"The reality is that OpenAI (and Hugging Face and others) have very little proper network expertise, so they just don't do security competently," Crowcroft added. "There's no evidence that this was anything relating to artificial superintelligence or the singularity — the logs and analysis from Anthropic just show a very tedious pile of script kiddie automation, which resulted in the OpenAI system getting at some data but no confidential stuff whatsoever."

Assessing intelligence

British computer scientist Alan Turing first described the "imitation game" in his seminal 1950 paper on machine intelligence. (Image credit: Pictures from History via Getty Images) To measure whether these systems are actually gaining true intelligence, researchers have historically relied on standardized benchmarks. One of the earliest examples was the " Turing test ," devised by computing pioneer Alan Turing , which evaluates whether an AI could convincingly fool an evaluator into believing it was human.

Researchers have claimed for years that AI systems can reliably pass this test , but Anil Seth , a professor of cognitive and computational neuroscience at the University of Sussex in the U.K., said this exam "is a test of human gullibility rather than machine intelligence."

"It's a test of what it would take for a human to decide that an AI is intelligent, which is kind of the reason that it's been a bit of a moving benchmark, because what it takes to convince us changes," he told Live Science. "It's all about creating typed text on a screen, and that's a very limited window into what we mean by intelligence."

To measure AI's cognitive ability more objectively, researchers are developing new metrics. For example, the ARC-AGI test , developed by nonprofit consortium the ARC Prize Foundation, tests AI's ability to teach itself completely new skills in response to problems it hasn't encountered as part of its training, operating purely on visual input. On its most recent test on 24 July , the top-ranked AI model hit 30.2%, while humans generally score close to 100%.

Another advanced metric, Humanity's Last Exam, includes around 2,500 Ph.D.-level questions across a broad range of subjects and requires advanced reasoning capabilities. Although this benchmark is not strictly a test of AGI, experts say machines will be able to reliably ace this test as a prerequisite to meeting the definition.

The AI hype machine

While figures like Altman and Musk suggest that we've already crossed the point of no return on the path to AGI, others are more skeptical. Gary Marcus , a professor emeritus of psychology and neural science at New York University, argued in a recent blog post that "no matter how you slice it, we just are not actually there yet." Echoing Crowcroft's assessment of the Hugging Face attack, he said, "Had OpenAI ordinary guardrail classifiers been in place, it wouldn't have happened."

Crowcroft expressed skepticism of Silicon Valley's assertion that the singularity is upon us. "Musk is, like the folks at OpenAI and Anthropic, talking nonsense just to keep the hype afloat," he said. "The point at which AI improvements are being mainly achieved by using AI to code and optimize itself ... is certainly a thing slowly arriving. Mark Handley , a professor of networked systems at University College London, and OpenAI "rewrote their entire data center protocol stack using agentic programming fairly recently," Crowcroft added. "But using tools to make better tools is as old as the flint and iron age — and in computing, as old as compilers, debuggers and optimizers."

Marcus referenced mathematician I.J. Good's foundational 1960s paper on ASI . "The singularity is usually supposed to mean a step beyond AGI that can do anything a person ‪—‬ even an expert ‪—‬ can do, and much more," he said, calling the idea that AI has passed this point "laughable."

Marcus pointed to 10 tasks he devised with AI researcher Miles Brundage , executive director of the AI Verification and Evaluation Research Institute. AI should be able to do these tasks just as well or better than the best human experts to be classified as AGI. They include writing Oscar-caliber screenplays; drafting cogent, persuasive legal briefs without hallucinating any cases; and making Nobel-caliber scientific discoveries.

Scientist Ray Kurzweil has published two books on the topic of the singularity, including "The Singularity is Near" (Duckworth, 2005) and "The Singularity is Nearer" (Vintage, 2025) — in which he cut his timeline for achieving ASI. (Image credit: Big Event Media / Stringer via Getty Images) Crowcroft said misinformation is feeding the hype about the AI singularity. "I think the hype is a deliberate confusion with the human singularity idea — uploading consciousness from bio to silicon to achieve some sort of immortality — which is total gibberish right now," he said. "The other deliberate confusion is to conflate singularity with AGI, which is marginally less nonsense but still a long, long way off for lots of good technical reasons."

Related stories

Artificial general intelligence (AGI) may arise in 2027 with artificial 'super intelligence' sooner than we think

AI models will lie to you to achieve their goals — and it doesn't take much

MIT scientists have just figured out how to make the most popular AI image generators 30 times faster

Other experts — such as Emily M. Bender , a professor of linguistics at the University of Washington, and sociologist Alex Hanna , director of research at the Distributed AI Research Institute — have argued that the very idea of conscious machines is a tactic designed to promote commercial AI products.

Seth, meanwhile, argues that we'll only be able to identify the singularity in hindsight. "From anywhere you are on an exponential curve, things will always look impossibly steep in front of you and irrelevantly flat behind you," he said. "It's a very bad idea to use as evidence the idea that we seem to be at a critical point ... because that's just a property of wherever you are on an exponential curve."

AI's competence varies widely among tasks, he noted, stressing that true AGI requires a model to be good at all cognitive tasks.

"They're very good at something, like coding or math proofs ‪—‬ but commonsense reasoning is really not so good, and doing things in the real world is not great," Seth said. "The singularity has got to be good at everything and change everything. Being really good at one or two ‪—‬ or even a large number of ‪—‬ things is not being in the singularity ."