In July 2026, Sam Altman sat down on the Relentless podcast and said something that most people either missed or dismissed: we are living through the singularity right now. Not approaching it. Not heading towards it. Inside it.

The reaction was oddly muted. Perhaps that is because the word "singularity" has been cheapened by decades of breathless futurism. Perhaps it is because Altman has a commercial interest in sounding bullish. Or perhaps — and this is the interpretation I find most honest — it is because the singularity, if it is happening, does not feel like the Hollywood version. There is no sky turning red. No sudden discontinuity. Just a compounding curve that most of us are riding without fully grasping the gradient.

Let me try to unpack what Altman actually said, what the numbers look like, and where I think the honest uncertainties lie.

What Altman Actually Said

The core claim is not vague. Altman introduced what he calls the "capability multiplier" — a concrete measurement of how much more capable AI systems are compared to a fixed baseline roughly 20 months prior. His number: 52x. Current-generation models can do 52 times as much useful work as the systems available in late 2024.

More striking than the absolute number is the rate of change. Altman claims capability is doubling every 13 weeks. That is a three-month doubling period. For comparison, Moore's Law — the transistor-density doubling that powered the computing revolution — had a doubling period of about 18 to 24 months. What Altman is describing is roughly six to eight times faster than the pace that transformed the semiconductor industry over five decades.

If that rate holds even approximately, it implies that by mid-2027, AI systems will be roughly 1,000 times more capable than they were in late 2024. By late 2028, something like a million times. These numbers stop being meaningful in any intuitive sense, which is rather the point.

Is There Independent Evidence for This?

The temptation with any Altman claim is to dismiss it as marketing. OpenAI crossed $25 billion in annualised revenue in March 2026. The man has every incentive to talk up the technology. So let us look at data that does not come from OpenAI's PR team.

OpenRouter, which acts as a routing layer for dozens of AI model providers, published usage data showing that token volume on its platform grew 24,000 times between August 2023 and August 2026. That is not a typo. Twenty-four thousand x in three years. The doubling period in their data is approximately 11 weeks — even faster than Altman's 13-week claim.

The prediction markets tell a more measured story. Manifold Markets, which has over 1,100 active predictors on AI-related questions, currently prices the probability of an AI passing an adversarial Turing test by 2035 at a level suggesting broad consensus it will happen within that window. Kalshi, the US-regulated prediction market, puts the probability of OpenAI specifically achieving AGI by 2030 at around 40 per cent. Metaculus, which uses a more structured forecasting methodology, currently estimates weakly general AI being announced by February 2028 and an adversarial Turing test being passed by April 2029.

These are not fringe bets. They represent the aggregated judgement of thousands of people who are putting real money or reputation on the line.

Then there are the capability demonstrations that are harder to dismiss. Anthropic's Mythos system found previously unknown vulnerabilities in every major operating system and browser it was tested against. Not by searching for known patterns, but by reasoning about code in ways that surprised even the security researchers evaluating it. That is not a chatbot party trick. It is a qualitative shift in what AI systems can do autonomously.

The Case for Skepticism

I want to be honest about where I think the "we are in the singularity" framing overreaches.

First, the 52x capability multiplier is Altman's number. He has not published the methodology behind it in sufficient detail for independent verification. We do not know exactly what is being measured, how tasks are weighted, or whether the metric captures real-world usefulness or benchmark performance.

Second, doubling curves do not continue indefinitely. Every exponential in the history of technology has eventually hit an S-curve — a period of rapid growth followed by a plateau. The question is not whether the current pace will slow, but when.

Third, there is a meaningful distinction between capability and deployment. Even if models are 52x more capable in a laboratory sense, the infrastructure, regulatory frameworks, and human workflows needed to deploy that capability lag significantly. A model that can diagnose rare diseases better than a specialist is not useful if the NHS cannot integrate it into clinical pathways.

What This Means for the UK

Britain is the world's third-largest AI market by investment, behind the US and China. We have genuine strengths: DeepMind remains headquartered in London, the Alan Turing Institute does world-class research, and our universities continue to produce exceptional machine learning talent. The UK government's AI Safety Institute has also given Britain a credible role in the governance conversation.

But there are structural problems. The UK's compute infrastructure is thin compared to what hyperscalers are building in the US. Our regulatory approach, while thoughtful, moves slowly. And the most consequential deployment decisions — how AI gets integrated into healthcare, education, financial services, defence — are being made by organisations that are, frankly, not ready.

The singularity framing matters here because it changes the urgency. If Altman is right that capability is doubling every 13 weeks, then the gap between "AI can do this in principle" and "AI is doing this in practice" is closing fast. The UK cannot afford a five-year regulatory cycle for a technology that is transforming quarterly.

That does not mean panic. It means taking the data seriously and acting on it. The countries and institutions that move first on deployment — not just research, not just safety papers, but actual integration into productive systems — will capture disproportionate value.

The Honest Middle Ground

Here is where I land, and I want to be transparent that this is a judgement call rather than a certainty.

I think Altman is broadly correct that the rate of capability improvement is historically unusual and that the word "singularity" is not unreasonable as a description of what is happening at the frontier. The compounding is real. The data from independent sources — OpenRouter, prediction markets, capability demonstrations — is consistent with a period of unusually rapid and accelerating progress.

I also think the popular image of the singularity as a sudden, dramatic inflection point is misleading. What we are experiencing looks more like a very steep gradient — steep enough that it changes the strategic calculus for every organisation and government, but not so vertical that it renders all prior planning obsolete overnight.

The most important thing to understand is that the pace itself is the story. Whether you call it the singularity or just "very fast progress," a 52x capability gain in 20 months with a 13-week doubling period is not business as usual. It demands a different kind of thinking from policymakers, business leaders, and frankly from all of us.

The question is not really "are we in the singularity?" The question is whether we are going to act like it.

Frequently Asked Questions

Did Sam Altman really say we are in the singularity?

Yes. During the Relentless podcast in July 2026, Altman stated explicitly that we are currently living through the singularity. He framed it in terms of a "capability multiplier" — AI systems being 52 times more capable than they were 20 months prior, with capability doubling roughly every 13 weeks.

What is the capability multiplier?

It is a metric Altman uses to describe how much more capable current AI systems are compared to a fixed baseline. A 52x multiplier means the current generation can accomplish 52 times as much useful work as systems from late 2024. The methodology behind this specific number has not been independently published or verified.

What do prediction markets say about AGI timelines?

As of mid-2026, Kalshi prices roughly a 40 per cent probability of OpenAI achieving AGI by 2030. Metaculus forecasts weakly general AI announced by February 2028 and an adversarial Turing test passed by April 2029. Manifold Markets, with over 1,100 predictors, broadly expects AI to pass an adversarial Turing test by 2035.

Is the AI singularity already happening?

The answer depends on your definition. If the singularity means a period of self-reinforcing, accelerating AI capability gains, then the data is consistent with Altman's claim. If it means a sudden, discontinuous transformation of civilisation, that has not happened. The reality appears to be a steep but continuous curve rather than a vertical wall.

What does this mean for the UK specifically?

Britain has strong AI research capabilities and a credible governance role through the AI Safety Institute. However, the UK lags on compute infrastructure and deployment speed. If capability is doubling every 13 weeks, the gap between what AI can do in principle and what UK institutions are actually deploying will become a significant competitive disadvantage unless deployment accelerates.