Sam Altman says humanity is already “in the singularity.” Elon Musk agrees, while Google DeepMind CEO Demis Hassabis says we are standing at its foothills. Behind the dramatic language is a development that is becoming increasingly difficult to ignore: AI systems are beginning to help build the next generation of AI.
For decades, the technological singularity belonged mostly to futurists and science fiction. The idea was that artificial intelligence would eventually become capable enough to improve itself, creating an accelerating feedback loop in which increasingly intelligent machines design even more capable successors. At some point, technological progress could become so fast that humans would struggle to predict — or control — what came next.
That hypothetical future is now being discussed in strikingly different terms by some of the people building the world’s most advanced AI systems. OpenAI CEO Sam Altman has said humanity is already entering the process. Elon Musk has publicly agreed. Google DeepMind CEO Demis Hassabis has described civilization as standing in the “foothills of the singularity.” Axios highlighted the convergence this week, noting that several leaders at the frontier of AI increasingly view the transition as something that may already be underway.
The claim is extraordinary, but the debate is no longer based entirely on speculation. There is now evidence for at least one important element of the singularity hypothesis: AI is increasingly being used to accelerate the development of AI itself.
That does not mean machines have become independently self-improving superintelligences. Humans still determine research goals, operate the infrastructure and decide which models are trained and deployed. But one of the feedback loops long associated with the singularity may be beginning to form.
Sam Altman says the “takeoff has started”
Altman has been unusually explicit about his view. In his essay The Gentle Singularity, the OpenAI CEO wrote that humanity is already “past the event horizon” and that “the takeoff has started.” Rather than imagining the singularity as one dramatic moment, Altman argues that increasingly powerful AI could become normal so gradually that society may barely recognize the transition while it is happening.
That distinction matters. The popular image of the singularity usually involves a sudden event in which humans build an artificial intelligence that becomes capable of autonomously making itself dramatically smarter. Altman’s version is much less cinematic and potentially much more gradual.
AI writes more software and helps researchers conduct experiments. Those researchers then use AI to develop better models, which in turn make the researchers more productive. The increased productivity helps produce another generation of AI, and the cycle repeats.
Altman has described today’s AI-assisted research as a primitive form of recursive self-improvement, while acknowledging that it is not the same as an AI system autonomously redesigning itself. More recently, he put his position even more directly during an appearance on the Relentless podcast, saying: “We are now, like, in the singularity.”
The significance of that statement is difficult to ignore. It comes from the CEO of one of the companies competing most aggressively to develop increasingly capable artificial intelligence.
Elon Musk agrees — and Demis Hassabis sees the “foothills”
Altman and Elon Musk famously disagree about many things, but apparently not this. After Altman’s comments triggered debate, Musk wrote on X that “We are in the Singularity,” reinforcing his own long-standing argument that AI has already become superhuman across a growing number of domains.
That does not necessarily mean Musk and Altman use exactly the same definition of singularity. Computers have been superhuman at arithmetic for decades, chess engines have surpassed the world’s best players for years, and modern AI systems can search, recall and synthesize information at a scale no individual person could realistically match. None of those achievements alone constitutes a technological singularity.
The more important question is whether AI is becoming capable of accelerating the development of AI itself. That is where the argument becomes considerably more consequential.
Google DeepMind CEO Demis Hassabis has been slightly more cautious in his language. At Google I/O in May, Hassabis described humanity as being “at the foothills of the singularity.” He later explained that he sees the singularity less as one identifiable future date and more as the beginning of a technological era characterized by rapidly advancing machine intelligence.
Hassabis specifically pointed toward developments in AI agents, software engineering, science and mathematics as evidence that something fundamental is changing. Rather than suggesting that a machine has suddenly become an autonomous superintelligence, his argument centers on AI systems becoming capable of performing increasingly complex intellectual work with less human involvement.
And there is now substantial evidence that this is already happening inside the companies developing frontier AI models.
AI is beginning to help build AI
Perhaps the clearest example comes from Anthropic. In June, the company published an analysis titled When AI Builds Itself, examining how Claude is increasingly involved in developing Anthropic’s own AI systems.
Anthropic is careful to emphasize that full recursive self-improvement has not been achieved and may never happen. Even so, the numbers illustrate how quickly the development process is changing.
As of May 2026, Anthropic said more than 80% of the code merged into its codebase was authored by Claude. The company also reported that its typical engineer was merging around eight times as much code per day in the second quarter of 2026 as in 2024, although Anthropic cautioned that lines of code significantly overstate actual productivity gains.
More importantly, AI’s role is extending beyond simply generating code. Anthropic says Claude can already take relatively underspecified engineering problems and determine how to solve them. On well-defined research experiments, the company says its systems can match or outperform skilled human researchers at execution.
That is creating a fundamentally different development cycle. Historically, humans researched AI and then built better AI systems. Increasingly, humans and AI are conducting that research together, with AI performing larger portions of the work required to develop its successors.
The theoretical next stage would be an AI system capable of choosing important research problems, conducting the experiments, interpreting the results and using those discoveries to develop a better version of itself. That complete loop would represent something much closer to genuine recursive self-improvement.
Anthropic says we are not there yet. But the gap between AI-assisted AI development and autonomous AI development has suddenly become one of the most important questions in technology.
The missing capability may be judgment
Anthropic’s own research identifies a major limitation in current systems: AI is increasingly capable of executing experiments, but it remains much weaker at deciding which experiments are worth conducting.
That distinction separates technical execution from genuine research judgment. A powerful AI agent may be able to receive an objective, write the necessary code, run thousands of experiments, analyze the results and generate a detailed report. A leading scientist does something different: they identify which problems matter, recognize unexpected results, challenge existing assumptions and invent entirely new research directions.
Anthropic says substantial gaps remain when models are required to exercise this kind of goal-setting judgment. In practical terms, that gap may be one of the main barriers separating today’s AI systems from machines capable of autonomously designing their own successors.
That is an important limitation because humans remain deeply embedded in today’s development loop. People choose the objectives, control the computing infrastructure, verify results and ultimately decide whether a new model should be trained or deployed.
Critics say this still isn’t the singularity
Not everyone working in AI accepts the increasingly dramatic language coming from Silicon Valley. William Tunstall-Pedoe, a co-creator of Amazon Alexa and founder of AI company UnlikelyAI, has argued that today’s systems lack the kind of autonomous innovation that a true singularity would require. He describes the missing capability as “discovery intelligence” — the ability to independently generate genuinely important new ideas and reliably determine which of those ideas are valuable.
Modern language models can generate enormous numbers of possible answers, hypotheses and approaches. That does not necessarily mean they can recognize the scientific breakthrough hidden among thousands of plausible but unimportant ideas.
Tunstall-Pedoe therefore rejects the claim that the singularity has already arrived. His argument is that genuine runaway self-improvement would produce effects across science, technology and society that would be extraordinarily difficult to miss.
Other researchers make a similar distinction. Even when AI systems demonstrate surprising autonomous behavior, that does not necessarily qualify as a singularity if humans continue to set their objectives and retain the ability to stop them. By that stricter definition, humanity clearly has not crossed the threshold yet.
The definition of “singularity” may be changing
The disagreement may ultimately come down to what the word itself means. The traditional version of the singularity looks something like this:
Human-level AI → autonomous self-improvement → intelligence explosion → superintelligence.
Under that definition, current AI systems have not reached the singularity. They remain dependent on human-defined objectives, infrastructure and oversight. But a different interpretation is emerging among some AI executives. Under this version, the sequence looks more like:
AI assists humans → humans build better AI faster → better AI accelerates AI research → increasing portions of that process become automated.
Viewed that way, the singularity is not necessarily a single event. It is a process in which the feedback loop between AI capability and AI development becomes progressively faster.
Hassabis’ “foothills” metaphor and Altman’s “gentle singularity” both fit this interpretation. The transformation happens incrementally, even as the underlying rate of technological progress accelerates.
That version sounds less dramatic than a machine suddenly becoming conscious. It may also be more realistic.
The most important feedback loop in technology
Consider what happens if AI makes AI researchers twice as productive. Those researchers can build a better model faster, and the improved model can then make them even more productive. It might automate experiments that previously required weeks or months, allowing researchers to test vastly more ideas and produce another generation of AI sooner.
As the cycle continues, humans could spend progressively less time performing the research itself and more time choosing objectives and validating machine-generated discoveries. If AI eventually becomes capable of reliably choosing valuable research directions as well, one of the most important remaining human bottlenecks could disappear.
Nobody has demonstrated that complete loop. But frontier AI companies are already using artificial intelligence to automate growing portions of their own development processes, while simultaneously researching what could happen if those systems become capable of more autonomous improvement.
That is why the language coming from AI leaders matters. Talk of “the singularity” can easily sound like Silicon Valley hype, and some of it may indeed be designed to generate excitement around technologies developed by the very companies whose executives are making these predictions.
But the underlying development is measurable. AI systems are writing substantial amounts of software inside AI companies, running experiments, performing increasingly long autonomous tasks and helping researchers develop better AI systems.
The feedback loop is not closed. It is, however, no longer entirely theoretical.
Maybe the singularity won’t look like science fiction
The strangest possibility is that Altman’s gradual interpretation turns out to be closer to reality than the classic science-fiction version.
There may never be a morning when humanity wakes up to a headline announcing that the singularity has arrived. Instead, each individual development could appear relatively manageable. An AI writes code, then begins running experiments, then discovers something useful, then helps train another AI. That system becomes a better research assistant, allowing the next generation to arrive even faster.
Each step looks incremental when viewed in isolation. Only later does the rate of change become obvious. Whether that process deserves to be called the singularity remains fiercely disputed. What has undeniably changed is the way some of the world’s most influential AI leaders are talking about it.
They are no longer describing the singularity exclusively as a distant theoretical future. Some believe humanity is already experiencing its earliest stages. And if the singularity arrives gradually rather than all at once, the most difficult moment to identify may be the moment when it has already begun.
