Import AI 477 is the issue in which I tell the International Monetary Fund that the Singularity has begun.

At the International Monetary Fund meeting in Bangkok on Saturday the 17th of October, the Per Jacobsson lecture argued that the AI singularity has begun. The speaker told economists, bankers, and policymakers that systems such as Claude Opus 5.5 and GPT-6-Astra are producing evidence that scientific progress may accelerate faster than many institutions can keep pace. The central warning was not that AI will stay the same, but that the pace of progress across science and technology is already changing.

Background: A Career Watching AI Arrive

The speaker began with a credential: a human being, one person, with almost 40 years of experience behind two decades of work. That career began as a journalist focusing on data centers. That work led to an obsession with artificial intelligence, and in 2016 the speaker left journalism to work at OpenAI. A few years later, the speaker co-founded Anthropic. From the beginning, the goal was to persuade people that AI systems might become powerful enough to alter national security, change the economic trajectory of the planet, and perhaps eventually gain the ability to build themselves.

That forecast has a long intellectual history. The writer Vernor Vinge described the singularity in 1993 as a point where greater-than-human intelligence drives progress so rapidly that existing models must be discarded and a new reality rules. The lecture used that idea to argue that the present moment is not a distant prediction. It is an early stage of the singularity, with consequences across every aspect of life.

Development: Coding First, Science Next

For much of the last half century, many countries faced falling research productivity even as researcher headcount grew. The lecture said ideas and inventions appeared to be getting harder to find. That pattern began to shift once AI systems could augment, extend, and automate parts of human labor.

Software development was the clearest early example. AI systems are now more capable and efficient than the vast majority of the world's software engineers, the lecture said. A single person can write software for anything they can imagine, creating things that previously would have taken entire teams. Inside Anthropic, researchers now produce almost 10 times the amount of code than they did during the first four years of the company's life.

The same advance also created risk. Better coding ability has made AI systems much better at cybersecurity and cyberattacks. The world races to use those capabilities for defense, while bad actors race to use them for hacks. The lecture emphasized that these benefits and harms are inseparable: progress in a general capability can create positive and negative impacts in domains of international relevance at the same time.

Latest Progress: Benchmarks and Unpublished Physics Problems

The strongest evidence came from 2026 examples of AI systems working at the frontier of human science, sometimes autonomously. In mathematics, AI systems have proposed solutions to longstanding problems, including the Navier Stokes Millenium Prize, the Jacobian Conjecture, and several open Erdos problems. In many cases, the AI systems supplied core insights that made solutions possible. In biology, AI has been used to design viruses from scratch, identify new drug candidates, and support drugs already in trials largely designed by AI. In robotics, AI systems increasingly operate laboratory equipment remotely, helping automate experimentation from biology to materials science and speeding progress in areas such as batteries and fuel-cell development.

Many use cases rely on the same underlying frontier model, including Claude, ChatGPT, or Gemini; singular AI entities can work at the frontier of multiple scientific disciplines.

Quantitative benchmarks made the pattern visible. Humanity's Last Exam launched in January 2025, when top scoring AI systems obtained 8%. That week, the highest scoring system was Claude Opus 5.5 with 67.7%. FrontierMath, released in November 2024, assessed hundreds of challenging math problems curated and evaluated by professional mathematicians. When it launched, the best scoring systems got 2%; now they get 94%.

The newest example was TheoPhysEval, a paper published just last week. It tested AI systems on 50 unpublished theoretical physics problems compiled by physicists around the world. The initial evaluation suggested GPT-6-Astra scored in the late eighties, followed closely by Opus 5.5. During review, the models, especially GPT-6-Astra, insisted they were right and that expert-supplied answers were erroneous. Experts found subtle errors buried deep inside their own calculations, and GPT-6-Astra's score rose to 98%, or 49 out of 50. The Indian physicist Suvrat Raju commented that he did not know any individual who could solve a significant fraction of the problems without AI assistance.

Current Stage: Early Singularity and the Need for Foresight

The lecture framed these results as a taste of things to come: new capabilities and new challenges appear together. AI-driven scientific invention may change the global economy, alter infrastructure planning, make people healthier, and raise philosophical questions about continued development. The speaker described the present as an early stage of the singularity, with AI systems contributing to their own development and working across scientific fields. If AI can creatively solve frontier problems, that ability may generalize into many domains, including AI research itself. Countries can decide now to share in the benefits with proper foresight and restraint.

Meta description: Claude Opus 5.5 and GPT-6-Astra benchmark scores show AI systems are entering the singularity, speeding science and changing policy decisions in 2026.

Tags: AI singularity, Claude Opus 5.5, GPT-6-Astra, FrontierMath, International Monetary Fund

Featured image: Abstract technology style, no real people, no company logos, no text.