Has the singularity begun? Anthropic's Jack Clark says it has

What the singularity actually means

The technological singularity is the idea that AI could accelerate scientific and technological progress so fast that human society can no longer predict or keep pace with it.

The mechanism is a loop. An AI system helps build a more capable AI system, which helps build a more capable one again, and each turn of the loop arrives faster than the last.

Vernor Vinge set out the modern version in 1993. Ray Kurzweil later popularised it with a date attached, around 2045. For most of that time it has been an argument rather than an observation, and our sister site covers what the term requires and what it does not.

Clark's claim is that the loop has started turning.

His evidence, and it is more specific than the headlines

Clark is not claiming AI has become conscious, all-knowing or capable of running anything by itself. His argument is narrower: AI is now contributing to scientific discovery in ways that change how fast new knowledge arrives.

Mathematics

He points to AI systems proposing solutions to long-standing open problems, naming the Navier-Stokes Millennium Prize problem, the Jacobian Conjecture and several open Erdős problems, and says that in many cases the AI supplied the core insight rather than the arithmetic.

He also cites two benchmark movements. On Humanity's Last Exam, top scores went from around 8% at its January 2025 launch to 67.7% with Claude Opus 5.5. On FrontierMath, from 2% at launch to 94%.

The detail most worth pausing on is a third. On TheoPhysEval, a 50-question theoretical physics test published the week before the speech, GPT-6-Astra scored 49 out of 50. The physicist Suvrat Raju reported that when the models disputed the experts' answers, the experts went back and found subtle errors in their own calculations.

That last one is a different category of claim from a benchmark score. It is a system correcting specialists in their own field.

Proposing a solution is still not proving one, and AI-generated mathematics needs the same verification as anyone else's. But "it did the sums faster" does not describe what is being reported here.

Medicine and biology

Clark says AI has been used to design viruses from scratch, to identify new drug candidates, and that some drugs now in clinical trials were largely designed by AI.

Note the first of those. Designing viruses from scratch is not a throwaway line in a speech about economic growth, and Clark does not treat it as one - he singles out biological research as the area needing the strictest human oversight and the most transparency.

On the upside, faster discovery does not mean faster medicine. A compound still has to clear safety testing, clinical trials and manufacturing. Clark makes this point himself, and goes further: if discovery becomes abundant, drug testing and clinical translation become the bottleneck.

Automated experiments

AI systems are increasingly able to operate laboratory equipment remotely, which means planning, running and interpreting experiments rather than only suggesting them. Clark cites faster progress in batteries and fuel cells.

The constraint here is physical and does not go away. Laboratories need equipment, materials, energy, funding and people. Digital intelligence does not dissolve any of that - which is why Clark's ask of governments is scientific equipment and robotics capacity, not just more compute.

The part that makes it a singularity argument

Everything above describes AI accelerating science. The singularity claim needs one more step: AI accelerating the production of AI.

That is already happening in the weak sense. Models write software, analyse results, design experiments and support engineering, and some of that work goes directly into building the next generation of models.

A substantial paper published in September 2026, What If Automating AI R&D Triggers an Intelligence Explosion?, examined whether automating more of that process could produce a period of exceptionally fast capability growth. Its authors include Geoffrey Hinton, Yoshua Bengio, OpenAI's Jakub Pachocki and Clark himself.

Their conclusion was not that an intelligence explosion is coming. It was that the preliminary evidence makes the possibility serious enough to need attention now, and that governments should require disclosure and allow auditors in.

Keep the three claims apart, because almost all the confusion in this debate lives in the gaps between them. AI helps human researchers - established. AI automates substantial parts of AI research - stronger, and needs stronger evidence. AI repeatedly improves the systems that improve AI, compounding - that is the singularity, and it is not demonstrated. Our piece on AI helping build AI works through where the evidence currently stops.

The $44 trillion number

Clark told the room that Anthropic's economic modelling projects US GDP reaching $44 trillion by 2030, against just over $30 trillion now. That is roughly a 50% increase in four years, and it is the number that travelled furthest.

It is a scenario, not a forecast, and Clark says so. It assumes models can autonomously complete a broad range of tasks, that they spread quickly through the economy, and - he is explicit about this - that entirely new industries and inventions appear. Remove any one of those and the number does not hold.

The interesting question underneath it is not whether the figure is right. It is what happens to an economy when discovery, software and engineering all get substantially faster. New industries could appear. Costs could fall. None of which says anything about who ends up holding the gains, and rapid technological progress has no built-in mechanism for distributing them fairly.

Is this AGI arriving?

No, and the two ideas are worth keeping separate.

AGI describes a system capable across a broad range of intellectual tasks at or beyond human level. The singularity describes a period in which progress accelerates beyond our ability to predict it. They are related - AGI is one proposed trigger - but one is a property of a system and the other is a property of a period in history.

You can argue for acceleration without anyone agreeing a definition of AGI has been met, which is what Clark is doing. There is also no agreed test for when a singularity has started, so his claim cannot be checked the way a benchmark score can. It is an interpretation of current evidence, offered by someone unusually close to it.

The definitions themselves are the problem as much as the evidence is - the main definitions in use disagree about what would settle the question.

The part of the speech that got least attention

The headline was the singularity. The more unusual content was what Clark asked for.

He called for governments to keep building electricity, data centre capacity and AI supply chains, and then to move on to scientific equipment and robotics. He said each scientific discipline will have to decide which experiments humans run and which AI runs, and how much humans need to understand discoveries that AI produced. He said labs including his own must decide how much oversight to apply to AI-driven AI research, and that rules will probably be needed.

And he argued for restraint - for deliberately slowing the development of the technology itself.

That is a co-founder of a frontier lab, speaking at the IMF, asking for his own industry to be constrained. Whatever you make of the singularity claim, that is the sentence worth remembering.

Underneath it sits a principle this site keeps returning to: the fact that a system can do something is not an argument that it should be allowed to. Capability is not permission.

So has it begun?

We do not know, and anyone who tells you otherwise is selling an interpretation as a finding.

What the evidence supports: AI systems are contributing to serious intellectual and scientific work, measurably better than two years ago, in ways that occasionally correct human specialists.

What it does not establish: that the rate continues, that contribution compounds into self-acceleration, or that discoveries convert into economic growth and better lives on any particular timetable.

Clark's position deserves weight because he is close to the systems and has identified a real shift in what they contribute. Proximity is also the reason to read him carefully. He is describing an industry he helps run, using his own company's models as evidence and his own company's economics as the projection.

Treating that as either science fiction or settled fact would both be mistakes. The useful posture is the one this site tries to hold: watch what is actually demonstrated, keep the categories apart, and notice when a measured result turns into a forecast somewhere between the paper and the headline.

We may look back at 2026 as the year the loop started turning. We may look back at it as a year of real but ordinary progress that got narrated as something larger. The evidence will decide it, and not for a while yet.

Sources

• Jack Clark: Import AI 477: In which I tell the IMF the Singularity has begun (19/10/26)
• CNBC: AI rapidly improving, could vastly boost US GDP, Anthropic's Clark tells policymakers (17/10/26)
• Centre for the Governance of AI: What If Automating AI R&D Triggers an Intelligence Explosion? (September 2026)