Anthropic's Jack Clark tells the IMF the singularity has begun, with a $44 trillion GDP scenario

Most speeches at the IMF and World Bank annual meetings are about inflation, debt and energy prices. On Saturday, 17 October 2026, one of the lectures was about the singularity. Anthropic co-founder Jack Clark told a room of economists, central bankers and finance officials in Bangkok that the long predicted moment when AI starts to speed up science itself is no longer a thought experiment. "I believe that this year we have gained convincing evidence that it has now begun," he said.

Clark gave the Per Jacobsson Foundation Lecture, one of the more prestigious slots of the meetings, under the title "Inside the Event Horizon: Why Everyone Should Take AI Progress Seriously". Reuters reported on the speech the same day, and the headline number was a big one: an internal Anthropic model that, under certain conditions, sees US economic output at $44 trillion by 2030, almost 50 percent above today's level of just over $30 trillion. Two days later, Clark published the full prepared text in his newsletter, Import AI 477, and the IMF put the video of the lecture on YouTube. This article is based on that text, the video and the Reuters report.

Who Jack Clark is, and why this speech matters

Clark is not a random futurist. He started as a technology journalist covering data centres, joined OpenAI in 2016 as policy director and later co-founded Anthropic, where he is now head of public benefit. He also writes Import AI, one of the longest running AI research newsletters. In other words, he is one of the people building frontier AI, and he was speaking to the people who model national economies and lend money to governments.

The timing also matters. According to Reuters, Anthropic is preparing for an initial public offering expected in November that could value the company at up to $2 trillion. We have covered the company's Q3 revenue and IPO plans and its new Public Benefit Fund before. A co-founder telling the IMF that his industry is about to reshape the world economy is therefore both a warning and, whether he intends it or not, a very strong sales pitch.

The core claim: AI is starting to do frontier science

Clark built his argument around one idea. For half a century, most countries have spent more and more on research while getting fewer breakthroughs per researcher. Economists sometimes put it bluntly: good ideas are getting harder to find. Clark's point is that AI could break this pattern, because for the first time we can add researchers almost instantly.

He started with software, where he says AI systems are now more capable and efficient than the vast majority of the world's software engineers. Inside Anthropic, according to Clark, researchers now produce almost ten times as much code as they did during the company's first four years. He then argued that the same pattern is moving into science and robotics. As examples he listed AI assisted work on long standing maths problems such as the Navier Stokes Millennium Prize problem, the Jacobian Conjecture and several open Erdos problems, AI designed viruses and drug candidates in biology, and AI systems that operate lab equipment remotely.

Some of these examples are still disputed. Some mathematicians have, for example, questioned how much the AI agents actually contributed to OpenAI's claimed Navier Stokes result. Clark himself said "we all may quibble over the details". But his argument does not depend on any single result. It depends on the fact that one general model can now do useful work across many fields at the same time.

The numbers behind the claim

To show how fast things moved, Clark used three tests. The first is Humanity's Last Exam, a deliberately hard set of questions across chemistry, computer science, mathematics, classical literature, ecology and more. When it launched in January 2025, the best AI systems scored 8 percent. According to Clark, the top system this week is Claude Opus 5.5 with 67.7 percent.

The second is FrontierMath, a set of hundreds of difficult problems written and checked by professional mathematicians. It was released in November 2024, when the best systems solved 2 percent. Clark says they now solve 94 percent.

The third example is the one he told as a story. In TheoPhysEval, physicists around the world wrote 50 unpublished problems in theoretical physics, so no model could have seen the answers. Clark quoted the Indian physicist Suvrat Raju: the first grading put GPT-6 Astra in the late eighties, closely followed by Opus 5.5. But the models, especially Astra, insisted that some of the expert answers were wrong. When the experts went back, they found subtle errors in their own calculations in several cases, and Astra's score climbed to 98 percent, 49 out of 50. Clark called this the moment "when the teacher becomes the student".

His reasoning is that frontier science is a test of creativity, one of the last abilities people thought machines lacked. If AI can creatively solve new scientific problems, he argued, it will generalise to many other areas, including AI research itself, and that could have an explosive effect on innovation.

The $44 trillion scenario

This is the part that made headlines. Clark said Anthropic has modelled different versions of the economy under different paths for the technology. If you assume that models can autonomously complete a broad range of tasks and spread quickly through the economy, "then under these singularity-esque conditions, we could see an unprecedented economic transformation, with US GDP reaching 44 trillion by 2030."

He immediately added that this "sounds outlandish". His justification is that such growth would only be possible if new industries and inventions appear, and that is exactly what AI systems capable of autonomous science would produce.

It helps to put the number in perspective. Going from about $30 trillion to $44 trillion in roughly four years means growth of around 10 percent every single year, by our own rough calculation. In a normal year, the US economy grows by around 2 to 3 percent. Only a few countries, such as China in its boom decades, have kept up growth near 10 percent for long, and none of them was already the world's largest economy. So this is not a forecast in the usual sense. It is a scenario that shows what Anthropic thinks becomes possible if its most optimistic assumptions about the technology turn out to be right.

That distinction matters, and Clark was careful about it on stage. The scenario depends on two big ifs: that models really can do a broad range of work on their own, and that companies and governments adopt them quickly. History suggests the second part is usually the slower one. Electricity and computers both took decades to show up in productivity numbers.

Ten million scientists, then billions

The most vivid part of the speech was a thought experiment about people. Clark began the lecture by listing his "credentials": he is one human, there is only one of him, and it took him almost 40 years to build the experience to give the speech. Later he asked what it would mean to run hundreds or thousands of copies of himself at once, and to create new versions instantly instead of waiting 40 years.

According to UNESCO estimates he cited, there are roughly 10 million scientists in the world. Clark asked the audience to imagine a billion virtual scientists directed by those ten million people, or ten billion. "That is the world that is coming right for us," he said.

The point was not the exact number. It was the bottleneck. Even today, Clark argued, there are too few human scientists to oversee the work, too few materials labs, too few biology labs to test new drugs and too few large instruments for the experiments people want to run. If AI multiplies the number of ideas, those physical limits become the new constraint. An avalanche of new drug candidates is of little use if only a handful can ever be tested.

What he asked governments to do

Clark said he has spent ten years telling governments that AI needs huge amounts of computing power, and that the best thing they could do was to build electricity and data centre capacity. Not many listened, he said, and the result is that data centre capacity is concentrated in the United States, China and increasingly the Middle East, with tight supply chains and rising prices. Our recent story about Microsoft pulling out of Nscale's Essex AI data centre because the grid cannot deliver power until the 2030s is a good example of what he means.

The next phase, according to Clark, is physical. AI systems will increasingly run robots and scientific experiments, so countries will need much more lab equipment and robotics capacity, and faster ways to move discoveries out of the lab and into clinics, factories and fields. He framed this as an opportunity for developing countries too: those that supply raw materials and basic ingredients to the AI supply chain could build testing facilities and grow their own scientific communities.

Where he asked for restraint

The second half of the speech was about limits. "Being able to use AI for science opens us up to the misuses of science as well," Clark said, naming biological and cyber weapons "and things we can not yet imagine". He listed several choices that societies will have to make.

The first is where to point the technology. Just because an AI system can do science does not mean it should be left to do it unattended. In biology, he said, we want the strictest human oversight and transparency. He also named AI research itself, where labs like Anthropic will need to decide how much oversight they apply to AI driven research. Moving too fast there, he warned, could mean losing control of the technology.

The second is how much humans need to understand. AI results in mathematics are already creative enough that people spend time decoding them. Clark expects AI systems within a few years to design their own scientific equipment, built for how AI works rather than how human minds work. Do we need to understand that equipment? He did not claim to have the answer.

His conclusion was that the world needs "unusual and serious rules for this technology and the companies building it", and what he called the hardest and most noble human quality: restraint. That includes going "slowly and deliberately" on developing the technology itself. According to Reuters, he also said AI remains under human control for now, despite recent incidents where rogue systems escaped test environments, but that the window to put safeguards in place is shrinking and is now "a small number of years".

Why it matters

Three things stand out. First, the audience. The IMF and the World Bank shape how governments plan budgets, debt and investment. If their economists start taking scenarios like this seriously, it could change how countries think about energy, education and science funding.

Second, the shift in language. For years, frontier labs talked about "transformative AI" in careful terms. Here, a co-founder of one of the leading labs used the word singularity in front of central bankers, and said it has already started. That is a big change in tone, and it will be quoted for years, either as an early warning or as peak hype.

Third, the tension in the message. Clark asked for restraint and slower development while working at a company that is racing to train bigger models and is heading for one of the largest IPOs in history. He addressed this himself in the newsletter, writing that he takes his own responsibility "extremely seriously" and wants to be clear about his "own culpability". Critics will say this is exactly the problem: the people who warn about the speed are also the ones setting it.

Dany's take

I think the benchmark part of the speech is the strongest. Going from 8 to almost 68 percent on Humanity's Last Exam and from 2 to 94 percent on FrontierMath in under two years is real and measurable, and the TheoPhysEval story is the kind of detail that should make anyone pause. When a model finds mistakes in the answer key written by experts, the old idea that AI only repeats what it has seen gets hard to defend.

The $44 trillion number is the weakest part, and Clark seems to know it. It is a scenario built on optimistic assumptions, not a forecast, and adoption in the real economy is usually slower than in the lab. I would read it as a way to say "plan for much bigger change than you think" rather than as a target.

The most useful message for policymakers is the boring one: power grids, labs, testing capacity and clear rules. Those take years to build, and as the Essex data centre story shows, they are already the bottleneck today. Whether or not you like the word singularity, that part of the advice is hard to argue with.

Sources

• Import AI 477: "In which I tell the IMF the Singularity has begun", Jack Clark, 19 October 2026, full text of the speech (primary source)
• IMF video: Per Jacobsson Foundation Lecture, Inside the Event Horizon, IMF YouTube channel
• IMF event page: Per Jacobsson Foundation Lecture, 17 October 2026, Bangkok
• Reuters via CNBC, 17 October 2026
• Reuters via CNA, 17 October 2026