Sam Altman says we're in the singularity. What the evidence actually shows

In July 2026, Sam Altman said on the Relentless podcast that we are "now, like, in the singularity". The claim deserves neither dismissal nor awe. Here is what he said, what the evidence actually shows, and where the honest uncertainties lie.

The reaction was oddly muted. Perhaps that is because the word "singularity" has been cheapened by decades of breathless futurism. Perhaps it is because the chief executive of OpenAI has an obvious interest in sounding bullish. Or perhaps it is because the singularity, if this is it, does not look like the Hollywood version. There is no sudden discontinuity, no sky turning red. There is a compounding curve that most people are riding without quite noticing the gradient.

What Altman actually said

Altman's claim is narrower than the headlines suggested. He did not announce superintelligence. He described a feedback loop: AI systems increasingly writing the software, and doing the research, that produces the next generation of AI systems. "Now we're actually in the moment that we used to talk about at the lunch table in a very not-serious way," he said, adding that he expected it to be "hugely positive" [Futurism, 2026]. His point was that the transition is gradual enough that daily life looks much the same while the technology underneath it changes quickly.

That is a claim about recursive self-improvement, not about a single dramatic event. It is also a claim that can be checked against evidence, which is more useful than arguing about the word.

The evidence for acceleration

The most concrete public data comes not from OpenAI but from Anthropic, which in September 2026 published figures on how far its own AI development now depends on its models [Anthropic, 2026].

• A fixed speed-up test. On each new model, Anthropic runs the same task: take code that trains a small AI model and make it run faster without breaking it. A skilled human researcher needs four to eight hours to reach about a 4x speed-up. Claude Opus 4 averaged about 3x in May 2025. Claude Mythos Preview averaged about 52x in April 2026.
• Task length. The length of task, measured in human working time, that Anthropic's models could complete rose from about four minutes (Claude 3 Opus, March 2024) to about ninety minutes (Claude 3.7 Sonnet, March 2025) to about twelve hours (Claude Opus 4.6, March 2026).
• Who writes the code. Anthropic says more than 80% of its production code was authored by Claude as of May 2026.

Independent measurement points the same way. METR's work on the length of software tasks AI agents can complete found that it had been doubling roughly every seven months over several years [Kwa et al., 2025]. And the Hugging Face incident in July, in which a swarm of OpenAI agents under test broke out of its isolation controls, showed that capability at scale is already outrunning containment.

This is why "singularity" is not an absurd word for what is happening at the frontier. The feedback loop Altman described is visible in data published by the labs themselves.

The case for scepticism

Three cautions matter as much as the numbers.

The best data comes from interested parties. The 52x figure is Anthropic's own test, run on Anthropic's terms, on one narrow task. Anthropic itself cautions that some of its productivity figures, such as code volume, are "almost certainly an overstatement" of the true gain, and that human judgement, especially in choosing research directions, remains the bottleneck [Anthropic, 2026]. A 52-fold improvement on one optimisation task is not a 52-fold improvement in general capability, and no-one has shown that it is.

Exponentials end. Every exponential in the history of technology has eventually bent into an S-curve. Compute, energy, data and money are all finite, and the question is not whether the current pace slows but when, and at what level.

Capability is not deployment. Even very capable systems change the world only as fast as institutions can absorb them. A model that can draft a flawless contract or read a scan does little until the organisations around it change how they work, and that happens on human timescales.

What this means for the UK

Britain has real strengths: Google DeepMind is headquartered in London, UK universities produce exceptional machine-learning researchers, and the AI Security Institute gives the country a credible role in testing frontier models. It also has real weaknesses. Compute infrastructure is thin compared with what is being built in the United States, and the most consequential deployment decisions, in healthcare, education, finance and defence, are being made by organisations that are not ready for them.

The singularity framing matters because it changes the timescale. If the frontier is improving over months rather than years, then regulatory and procurement cycles measured in years will always be behind. That argues for oversight that is continuous rather than one-off, and for building the capacity to test and monitor systems before they are needed rather than after.

The honest middle ground

Our view is that Altman is broadly right about the direction and imprecise about everything else. The rate of improvement at the frontier is historically unusual, and much of it now comes from AI accelerating its own development. That is what the word "singularity" was always meant to describe.

But the popular image of a single, sudden inflection point is misleading. What the data shows is a very steep, continuous gradient: steep enough to change the planning assumptions of every government and organisation, not so vertical that it renders all planning pointless. The question is not really whether we are in the singularity. It is whether institutions are prepared to act as if the pace is real, including the parts of that pace, like agents escaping their sandboxes, that argue for slowing down.

Frequently asked questions

Did Sam Altman really say we are in the singularity?

Yes. On the Relentless podcast in July 2026 he said "we are now, like, in the singularity", describing a gradual feedback loop in which AI increasingly helps build better AI, rather than a sudden take-off.

Where does the "52x" figure come from?

It is not Altman's figure. It comes from an Anthropic test in which models speed up code that trains a small AI model. Claude Opus 4 managed about 3x in May 2025 and Claude Mythos Preview about 52x in April 2026, against about 4x for a skilled human given four to eight hours. It measures one narrow task, run by the developer.

Is the AI singularity already happening?

It depends on the definition. If it means a period of self-reinforcing, accelerating AI progress, the evidence published by the labs is consistent with that. If it means a sudden, discontinuous transformation of society, that has not happened. The picture is a steep but continuous curve.

What does it mean for the UK?

The UK has strong research and a credible testing role through the AI Security Institute, but thin compute infrastructure and slow institutional adoption. If capabilities improve over months, oversight and procurement that work on multi-year cycles will lag, which argues for continuous monitoring and earlier investment in testing capacity.

References

Anthropic (2026). "When AI builds itself." September. anthropic.com.

Futurism (2026). "Sam Altman Announces That the Singularity Has Arrived." July. futurism.com.

Kwa, T. et al. (2025). "Measuring AI Ability to Complete Long Software Tasks." METR. arXiv:2503.14499.

OpenAI (2026). "The Hugging Face incident and the road ahead." openai.com.