Import AI 477: In which I tell the IMF the Singularity has begun

1. Key Themes

Theme: AI is now operating at the frontier of science, across disciplines, from a single underlying model

Frontier science results are arriving in math, biology, and lab automation

Clark argues 2026 delivered the first real evidence of AI doing frontier-level science, with concrete examples across fields.

"In mathematics, AI systems have been used to propose solutions to longstanding mathematical problems, like the Navier Stokes Millenium Prize, the Jacobian Conjecture, and several open Erdos problems. In many cases, AI systems were responsible for core insights that made the solution possible."

"In biology, researchers have used AI to design viruses from scratch, to identify new drug candidates for testing, and there are already drugs in trials today which have been in large part designed by AI."

"In robotics, AI systems are now increasingly able to autonomously operate laboratory equipment remotely, making it possible for us to further automate experimentation in domains from biology to materials science, making progress faster in areas like batteries and fuel-cell development."

The generality of one model is the real shift

The significance isn't any single result but that one model spans disciplines.

"But the profound fact is that they all relate to the same underlying technology - and for most of these use-cases, the same single underlying frontier model. We now have singular entities - AI systems like Claude or ChatGPT or Gemini - which can carry out work at the frontier of multiple scientific disciplines."

Benchmarks are saturating at extraordinary speed

"When the exam launched in January 2025, top scoring AI systems obtained a score of 8%. As of this week, the highest scoring system is Claude Opus 5.5 with 67.7%."

"FrontierMath... When it launched, the best scoring systems got 2%. Now, they get 94%."

"Pick your favorite part of science and go looking for a test and you'll find this pattern - sudden and steep increases in capability, everywhere you look."

Theme: AI addresses the "ideas are getting harder to find" problem by scaling research labor

Virtual scientists could multiply research capacity by orders of magnitude

Clark frames the economic mechanism: research productivity has been declining, and AI changes both the quantity and training time of researchers.

"For the last half century, most countries have dealt with falling research productivity while researcher headcount grows rapidly - ideas and inventions appear to be getting harder to find."

"What happens when instead of ~10 million scientists, we have a billion virtual scientists directed by our own ten million? Or ten billion? That is the world that is coming right for us."

"This means that in the coming years, every human scientist on this planet will become more important and more powerful, as they gain the ability to supervise hundreds or thousands of like-minded experts on any problem they find interesting."

Anthropic's internal data and economic modeling point to large gains

"We've seen concrete evidence of this inside Anthropic, where our researchers now produce almost 10 times the amount of code than they did for the first four years of the company's life."

"If you assume that models will be able to autonomously complete a broad range of tasks and will diffuse quickly in the economy, then under these singularity-esque conditions, we could see an unprecedented economic transformation, with US GDP reaching 44 trillion by 2030."

Theme: The next infrastructure buildout is physical-world science capacity (robots, labs, equipment), not just compute

Compute was phase one; labs and robotics are phase two

Clark says governments largely failed to heed the compute/power advice and are now facing concentration and price pressure, and he's issuing a new call.

"In today's world, data center capacity is concentrated in the United States, China, and increasingly the Middle East, and the supply chains are concentrated and are under such contention that prices are going up."

"The next phase of the AI revolution is about the expansion into the physical world, specifically by having AI systems increasingly operate robots, and to perform scientific experimentation, sometimes together."

"Even with only 10 million scientists, we have too few human scientists to oversee the work, too few laboratories to do materials science analysis, too few biological labs to develop and test new drugs, and too few large pieces of scientific equipment in space or on the earth for the experiments we want to run."

Downstream bottlenecks (clinical, manufacturing) become binding constraints

"If discovery becomes abundant, we need to avoid bottlenecking ourselves downstream - an avalanche of new proposed drugs is of little use if only a handful can ever be tested."

Developing countries can leapfrog via supply-chain and testing facilities

"If you are a developing country that supplies the basic ingredients into the AI and scientific enterprise, like raw materials, then you have an opportunity to leapfrog: you can both build scientific testing facilities that, paired with human scientists and digital AI scientists, will let you grow a scientific community."

Theme: Capability gains come bundled with dual-use risk and require restraint and governance

Coding gains simultaneously boosted cyber offense and defense

"The coding improvements have meant AI systems have got much better at cybersecurity and cyberattacks –– leading the world to race to deploy these powerful capabilities to create better cyberdefenses, while bad actors race to use these systems to carry out hacks."

"The important thing to notice here is that these things are bound up, inseparably."

Key choices: where to deploy AI-driven science, and how much to understand it

"One such use case is biological research, where we want the strictest human oversight and transparency to maintain safety."

"Another is AI research itself, where labs like Anthropic will need to make decisions about how much oversight we apply to AI-driven research... We will need to move extremely carefully here, because otherwise we could lose control of the technology."

"We are only a few short years from AI systems not only doing science, but inventing their own new scientific equipment built for the ways AI systems work, rather than human minds. Do we need to understand that equipment? Should we?"

2. Contrarian Perspectives

An AI lab co-founder calls for deliberate slowing of the technology itself while building infrastructure faster Most industry rhetoric pairs "build faster" with "develop faster." Clark splits the two: accelerate physical and scientific capacity, but pace the technology.

"We will need to build the infrastructure to take advantage of this technology, but we will also need to go slowly and deliberately on developing the technology itself, to pace or otherwise control its development."

He also pushes for binding rules on companies, including his own.

"We are going to need to work together to create unusual and serious rules for this technology and the companies building it, and to exercise the hardest and I think most noble quality of people - restraint."

The singularity is not a future event but has already begun, and policymakers should plan for faster change than current forecasts This goes beyond the common "AI is transformative" framing to a claim that the discontinuity is already underway.

"They call this the singularity. I believe that this year we have gained convincing evidence that it has now begun."

"What I am here to explain today is that AI is on a course to bring about even more change than what we are expecting and we should prepare accordingly."

Humans become more valuable, not less, in the near term as AI scales Against the labor-displacement narrative, Clark argues AI is a force multiplier for scientists, with human supervision as the scarce resource.

"In the coming years, every human scientist on this planet will become more important and more powerful."

"We have too few human scientists to oversee the work."

Experts' answer keys may be the error source, so human-curated benchmarks can understate AI In TheoPhysEval, the models disputed the expert-supplied answers and turned out to be right in multiple cases.

"The models, especially Astra, insisted that they were right and that it was the expert-supplied answers that were erroneous. When we went back to the experts it turned out, in multiple instances, that the experts found some subtle errors buried deep inside their calculations. As they revised their answers, Astra's score started to inch up and it finally ended up with 98% (49/50)."

3. Companies Identified

Anthropic

• Description: AI lab co-founded by the author.
• Why mentioned: Source of internal productivity evidence and economic modeling; author's own culpability and governance dilemmas.
• Quotes: "We've seen concrete evidence of this inside Anthropic, where our researchers now produce almost 10 times the amount of code than they did for the first four years of the company's life." / "We have done some of our own modeling here..."

OpenAI

• Description: AI lab where Clark worked from 2016.
• Why mentioned: Career background; also maker of ChatGPT and GPT-6-Astra.
• Quotes: "in 2016 I left journalism to work at OpenAI, then co-founded Anthropic a few years later." / "AI systems like Claude or ChatGPT or Gemini - which can carry out work at the frontier of multiple scientific disciplines."

Google (Gemini)

• Description: Developer of the Gemini model family.
• Why mentioned: Cited as an example of a general-purpose model capable of multi-discipline frontier science.
• Quotes: "singular entities - AI systems like Claude or ChatGPT or Gemini."

Claude Opus 5.5 (Anthropic)

• Description: Anthropic's frontier model.
• Why mentioned: Top scorer on Humanity's Last Exam and a close second on TheoPhysEval.
• Quotes: "As of this week, the highest scoring system is Claude Opus 5.5 with 67.7%." / "GPT-6-Astra's score was in the late eighties followed closely by Opus 5.5."

GPT-6-Astra (OpenAI)

• Description: Frontier model referenced in the physics benchmark.
• Why mentioned: Top performer on TheoPhysEval, reaching 98% after expert corrections.
• Quotes: "Astra's score started to inch up and it finally ended up with 98% (49/50)."

Humanity's Last Exam, FrontierMath, TheoPhysEval (benchmarks)

• Description: Scientific and math evaluations for AI systems.
• Why mentioned: Quantitative evidence of rapid capability growth.
• Quotes: "TheoPhysEval... tests out how well AI systems can solve unpublished problems in theoretical physics."

IMF

• Description: International Monetary Fund.
• Why mentioned: Host of the Per Jacobsson lecture; the audience of economists, bankers, and policymakers Clark is trying to move.
• Quotes: "I wanted to use my privileged position of giving the Per Jacobsson lecture to try and tell the assembled audience of economists, bankers, and policymakers, that they should expect significant, rapid change in the future."

UNESCO

• Description: UN scientific and cultural agency.
• Why mentioned: Source of the ~10 million scientists estimate underpinning the "virtual scientist" scaling argument.
• Quotes: "According to estimates by UNESCO, there are approximately ~10 million scientists in the world."

4. People Identified

Vernor Vinge

• Description: Science fiction author and originator of the singularity concept.
• Why mentioned: Anchors the singularity framing historically.
• Quotes: "When greater-than-human intelligence drives progress, that progress will be much more rapid... a point where our models must be discarded and a new reality rules."

Suvrat Raju

• Description: Indian physicist who commented on the TheoPhysEval results.
• Why mentioned: Provides expert testimony that AI found errors in human experts' answers.
• Quotes: "I don't know of any individual who can solve any significant fraction of our final collection of problems without AI assistance."

Carl Sagan

• Description: Astronomer and author of The Pale Blue Dot.
• Why mentioned: Closing frame for the need for wisdom and prudence on a compressed timeline.
• Quotes: "a species very like us, but with more of our strengths and fewer of our weaknesses… more confident, farseeing, capable, and prudent."

Jack Clark

• Description: Anthropic co-founder, former OpenAI policy lead, Import AI author.
• Why mentioned: Speaker; positions himself as both narrator and agent of AI's progress.
• Quotes: "I find myself both a narrator of what is happening in AI, a principle agent behind what is happening in AI, and a translator that goes between many communities."

5. Operating Insights

Treat AI as a way to instantiate "copies" of expertise and redesign workflows around supervision, not execution

"What would it mean to be able to run hundreds or thousands of copies of me simultaneously? What would it mean if new versions of me could be created instantly, rather than waiting 40 years?"

"We can rapidly augment humans in domains that show promise by generating large amounts of virtual scientists to work alongside them."

Build for the downstream bottleneck: validation, testing, and deployment will be the constraint when discovery is abundant

"An avalanche of new proposed drugs is of little use if only a handful can ever be tested. The policymakers and their colleagues in this room will need to help us find ways to accelerate our ability to move discoveries out of the lab and into the clinic, the factory, and the field."

Decide deliberately, domain by domain, what to delegate to AI and what to retain, and how much to understand

"What experiments should they spend time on as a person, and what experiments should they let the AI system do itself? This is a subtle question to which I don't know the answer, but it is a question each discipline of science will need to answer."

6. Overlooked Insights

Benchmark validity may flip: models may become the arbiters of correctness in expert domains The TheoPhysEval episode suggests evaluation pipelines relying on human ground truth will need to adapt, since the models flagged real errors in expert-supplied answers.

"Some of the world's scientific experts built a test out of problems an AI could never have possibly read the answer to. And when they tested the AI, not only did it score high, but it got a higher score by finding errors in the answers that the humans had worked on!"

AI-designed scientific instruments may be uninterpretable by humans within a few years This raises a governance and verification issue beyond model interpretability: the physical tools of science may be optimized for AI cognition rather than human understanding.

"We are only a few short years from AI systems not only doing science, but inventing their own new scientific equipment built for the ways AI systems work, rather than human minds."