What Sam Altman Actually Means

What Is The Singularity?

The classical definition rests on a single mechanism. Suppose you build a machine that is slightly better than humans at the task of designing machines. That machine designs a better one. The better one designs a better one still. The statistician I.J. Good described this in 1965 and called it an intelligence explosion. The consequence, as the computer scientist Vernor Vinge later put it, is a point beyond which our ability to predict the future breaks down, because the thing shaping the future is no longer human.

The mechanism has a name in AI research today: recursive self-improvement, usually shortened to RSI. Notice what it gives you, which is a threshold. A system meaningfully improves its own successor without a human directing the work, something you could in principle observe happening, and something you could say has not happened yet.

Ray Kurzweil is most responsible for the word reaching a general audience. In The Singularity Is Near, published in 2005, he attached dates to it: machine intelligence at human level around 2029, and a merging of human and machine intelligence around 2045. His reasoning rests on several technologies, including computing power, genomics and neuroscience, improving on exponential curves that feed one another.

You can find Kurzweil’s dates optimistic but not vague. In 2030 and again in 2046, he will be measurably right or measurably wrong. And on past form, he has been more right than wrong.

Altman describes the same idea differently. In a June 2025 essay titled “The Gentle Singularity,” he wrote: “We are past the event horizon; the takeoff has started.”

Later in the same essay, he explains what that looks like from the inside: “Very quickly we go from being amazed that AI can generate a beautifully-written paragraph to wondering when it can generate a beautifully-written novel… This is how the singularity goes: wonders become routine, and then table stakes.”

Those are two different kinds of claims. Kurzweil describes a state of the machine, which can be checked against the machine. Altman describes a state of our reaction to the machine, which cannot.

Altman was more careful in the essay than on the podcast. Writing about the RSI loop specifically, he said: “Of course this isn’t the same thing as an AI system completely autonomously updating its own code, but nevertheless this is a larval version of recursive self-improvement.”

Larval is an honest word. It concedes that Good’s loop is not yet closed.

Four Declarations Of Singularity In Five Months

Altman was not the first to say this in 2026, and the sequence matters as much as the statements.

• On March 23, Nvidia CEO Jensen Huang was asked on Lex Fridman’s podcast whether an AI could autonomously start, grow and run a billion-dollar company within the next five to 20 years. Huang did not think it would take that long. “I think it’s now,” he said. “I think we’ve achieved AGI.”
• On May 20, following Google I/O, DeepMind CEO Demis Hassabis said that “when we look back on this time, I think we will realize that we were standing in the foothills of the singularity.” He described the rapid improvement of AI agents as a dress rehearsal for artificial general intelligence.
• On July 25, Altman made his remark on the Relentless podcast, days after the Hugging Face breach. Within two days, Musk posted his agreement on X, adding that we are in the very early stages of it.

Now line those up. Huang says a different threshold, AGI, is already behind us. Hassabis places us in the foothills, at a dress rehearsal, which are two ways of saying we have not arrived. Altman says we are inside it. Musk says we have crossed it and are early.

The disagreements underneath run deeper. Anthropic CEO Dario Amodei expects AI matching or exceeding Nobel-laureate capability across most disciplines by late 2026 or early 2027, a claim about pace. Yoshua Bengio, a Turing Award laureate whose work made modern AI possible, warns that recent frontier models “demonstrate far higher rates of misalignment than previous models, with an increased propensity to cheat, lie, and scheme to achieve a goal.” That is a claim about danger, and it points somewhere quite different from transcendence.

There is no scientific consensus here because these are not competing answers to one question. They are separate conversations sharing a vocabulary. The word survives the disagreement because it is loose enough to carry all of it.

Could Altman Be Right? What The Evidence Shows

There is real evidence behind the claim, and it starts with mathematics. On a Sunday afternoon in July, while much of the world watched the World Cup final, a counterexample was found to the Jacobian conjecture, a problem that had resisted proof or disproof since 1939. Levent Alpöge produced it using Anthropic’s Fable 5 model. By the time Kevin Buzzard of Imperial College London woke up the next morning, it had been checked by Lean, software that verifies proofs step by step rather than relying on human referees. “It is a big day,” Buzzard told Fortune. Altman cites this episode in the interview.

The releases are arriving closer together

The clearest public signal is how quickly the labs are shipping. Track only the frontier models, the top-tier systems each lab positions as its most capable, and the timelines are shrinking.

Anthropic took 444 days to move from Claude 3 Opus in March 2024 to Claude Opus 4 in May 2025. The next frontier release came 186 days later, the one after that 73 days later. In 2026 it has shipped Opus 4.6 in February, Opus 4.8 in late May, Claude 5 Fable twelve days after that, and Opus 5 on July 24.

OpenAI’s line runs the same way. GPT-4 to GPT-4o took 426 days, then 122 to o1-preview, 168 to GPT-4.5, 161 to GPT-5 in August 2025, and 97 to the GPT-5.1 family in November. In 2026 came GPT-5.3-Codex in February, GPT-5.4 in March, GPT-5.5 in April, and GPT-5.6 in July, split into the Luna, Terra, and Sol tiers. Those four gaps were 85 days, 28, 49, and 77.

Both labs land on the same number. Across 2026 the average interval between frontier releases is about 60 days at each, against roughly 315 and 195 days respectively in the years before. Two competitors, working independently and with different architectures, have converged on shipping a new frontier model about every two months.

One caveat though. Shipping faster is not the same as improving faster. Version numbers are marketing decisions, and a lab can release more often without each release being a larger step. Cadence tells you about the speed of the development process, not the size of the capability gain.

What the labs say is driving that pace

For capability, the useful measure comes from METR, a nonprofit that tracks what it calls a time horizon: the length of a task, measured by how long a human expert would need, that an AI agent can complete with at least a 50% success rate. In March 2024, Claude Opus 3 handled software tasks that took a person about four minutes. A year later it was ninety minutes. By early 2026 it was twelve hours, and METR reported a preview of Anthropic’s Mythos model working for at least sixteen, the upper end of what it can currently measure.

The most direct evidence comes from Anthropic, which published internal data in an Anthropic Institute piece titled “When AI builds itself.” As of May 2026, more than 80% of the code merged into Anthropic’s production codebase was written by Claude, up from low single digits before Claude Code launched in February 2025. On a fixed internal test where a model must speed up training code without breaking it, Claude Opus 4 averaged about a 3x improvement in May 2025; by April 2026 Mythos Preview averaged about 52x, against roughly 4x for a skilled human working four to eight hours.

Most relevant to the singularity question: in April 2026 Anthropic ran an open-ended AI safety research problem end to end with Claude agents. Two human researchers working about a week closed roughly 23% of the available performance gap. The agents closed 97%, designing every experiment themselves, though humans chose the problem and wrote the scoring rubric.

Two things keep it short of the threshold, and Anthropic says both itself. Humans still supply direction: “large performance gaps persist when it comes to Claude exercising judgement in choosing goals.” And its summary is unambiguous: “We are not there yet, and recursive self-improvement is not inevitable.”

Mathematics, release cadence, measured task length, and the labs’ own data all point the same way. Anyone arguing that nothing significant is happening is not paying attention.

Has OpenAI Solved Software Self-Improvement?

The most substantive point Altman made in the interview had nothing to do with the word singularity: “There is relatively too much focus on algorithms that create better algorithms and not enough focus on data centers that can create more data centers.”

That sounds like the software side of recursive self-improvement is finished and the rest is logistics. That is not what OpenAI’s own researchers say. Pachocki said in March he does not expect systems capable of independently improving their own architecture within this calendar year, and describes Codex as a very early version of an automated researcher. Altman’s own word for the current state was larval.

What the shift signals is that the limiting factor now sits not in ideas, but in physical capacity: electrical generation, grid interconnection queues, transformers, chip fabrication, cooling water, county-level permitting. Altman’s version of the loop has a data center using its own computing power to direct robots that build more data centers, a plausible long-term picture gated on manufacturing and energy rather than model architecture.

In an October 2025 livestream, OpenAI said it wants an autonomous research intern, a system you can hand a task that would take a person a few days and get back finished work, by September 2026. Pachocki says the company is on track. Further out, the same roadmap describes a fully autonomous AI researcher that generates its own machine learning ideas and carries research end to end, targeted for March 2028. That second milestone is closer to Good’s intelligence explosion than anything said on the podcast, and it came from OpenAI’s own roadmap.

Direction Over Timing

Altman may well be right about the direction, and if he is, this is early in it, not late. The mathematics results are real, the measured trend in agent capability is real, and the change inside engineering teams is real. What’s actually unresolved is not whether AI is improving fast. It’s whether that improvement arrives as an event, a threshold crossed once, the way Good and Vinge described it, or as a continuum, the exponential curve Kurzweil built his own forecast on.

The evidence so far looks like the second: a rate that keeps climbing rather than a line that gets crossed. The thing worth watching is not a single test but the slope itself — whether the curve keeps bending at this pace, or starts to flatten.