When the Student Corrects the Answer Key: Living and Working in the Active Singularity

Last week, a quiet tremor ran through the foundations of theoretical physics.

A group of researchers published a new benchmark called TheoryPhysEval (formally released on arXiv as TheoPhysEval), a collection of 50 formidable problems compiled by physicists around the world. Unlike standard AI benchmarks that scrape textbooks or the open internet, the answers to these 50 questions had never been published. They sat solely in the private research notes of the scientists who worked them out. As one of the paper’s authors, Indian physicist Suvrat Raju, noted, there wasn’t a single human alive who could solve any significant fraction of the collection without assistance.

When the researchers ran the latest frontier models against the test, Claude Opus 5.5 and GPT-6 Astra scored in the late 80s. That alone would have been a milestone.

But when the evaluation team looked closer at the problems the models “missed,” something extraordinary happened: the models—particularly Astra—insisted they were right, and that the human answer keys were wrong. When the researchers took the AI’s work back to the physicists who had written the problems, the scientists discovered subtle errors buried deep inside their own original calculations. As the human experts corrected their own math, Astra’s final score ticked up to 98%—49 out of 50 correct.

In a recent keynote speech at the IMF Annual Meetings in Bangkok, Anthropic co-founder Jack Clark pointed to this exact moment as a real-life manifestation of a classic narrative trope: two-thirds of the way through the film, the student officially becomes the teacher.

More plainly, Clark stood at the podium and delivered a message that many of us working alongside these systems have felt building for months: “The singularity is here.”

Discarding the Old Models

For decades, the word “Singularity” has carried the baggage of science fiction rapture—an abstract, apocalyptic, or utopian horizon where machines wake up and humanity becomes obsolete.

Clark strips away the mysticism and returns to writer Vernor Vinge’s grounded 1993 definition: the Singularity is simply the point where greater-than-human intelligences drive progress so rapidly that our existing models of the world must be discarded, and a new reality rules.

You can see the breaking of our old models in the raw numbers Clark brought with him. When the benchmark FrontierMath—a curated test of hundreds of grueling mathematical problems—launched in November 2024, the world’s best AI systems scored 2%. Today, they score 94%. When Humanity’s Last Exam launched in January 2025 to test frontier capabilities across chemistry, computer science, classical literature, and ecology, top models managed an 8%. Today, Claude Opus 5.5 sits at 67.7%.

Across biology, mathematics, and materials science, the curve is bending vertically. AI systems are proposing insights for the Navier-Stokes Millennium Prize problem, designing novel viruses from scratch, and guiding drugs currently in clinical trials.

And the most profound detail is that these breakthroughs aren’t coming from narrow, specialized calculators. They are emerging from the same underlying general models. We are no longer just building software tools; we have instantiated virtual polymaths.

One of Me, Ten Thousand of Them

To understand what this means for those of us navigating this shift in our daily work, look at how Clark introduced himself at the start of his talk: “I am a human being. There’s one of me, and it’s taken me almost 40 years to build the experience to give this speech.”

For the last half-century, human progress has been trapped in a quiet stagnation. Even as we put more researchers into the field, researcher productivity has fallen. Good ideas have simply gotten harder and more expensive to find. Training a single human specialist takes 10 to 15 years of targeted education and millions of dollars, and at the end of that pipeline, you still only have one human mind in one room.

Now, ask the question that is reshaping the global economy: What happens when you don’t have to wait 40 years? What happens when a single human researcher can spin up a thousand specialized digital polymaths instantaneously, cross-pollinating biology with materials science and mathematics in seconds?

Inside Anthropic, researchers are already producing nearly 10 times the code they did during the company’s first four years. A single person can now architect and deploy software that used to require an entire engineering department months of labor.

For the thoughtful professional, this changes the nature of human work from the ground up. If you are a writer, a coder, a scientist, or an entrepreneur, your value is rapidly shifting away from your ability to lay every brick by hand. Every human specialist is becoming a force-multiplied Orchestrator—a steward supervising fleets of virtual domain experts.

Where Digital Speed Hits a Physical Wall

If Anthropic’s economics team is right, this explosion of autonomous scientific inquiry could push US GDP to an astonishing $44 trillion by 2030—a projection that requires the invention of entirely new general-purpose technologies and industries.

Yet if you look outside your window, the world doesn’t look like a $44 trillion utopia just yet. Why? Because while intelligence has gone weightless, reality is still made of atoms.

For the past ten years, the primary bottleneck to AI progress was digital infrastructure: data centers, chip supply chains, and electric grids. Clark warns that while compute still matters, the next hard bottleneck is physical experimentation.

According to UNESCO, there are roughly 10 million active human scientists on Earth today. Even at that human scale, we suffer from acute shortages of wet labs for therapeutic validation and facilities for materials testing. What happens over the next thirty-six months when those 10 million human scientists begin directing 10 billion virtual researchers?

A fleet of AI polymaths can generate ten thousand promising battery chemistries or life-saving drug candidates before lunch, but those discoveries mean nothing if they are trapped in a server rack. They have to be synthesized in a physical lab, tested by physical robotics, validated through clinical trials, and moved through regulatory pipelines that were built for a 20th-century pace.

This is a massive wake-up call for the builders of the physical world. The next phase of the AI revolution won’t just belong to prompt engineers in Silicon Valley; it will belong to the communities, tradespeople, engineers, and developing nations that build the automated laboratories, wire the power grids, and construct the physical testing grounds where digital hypotheses meet physical matter.

The Interpretability Dilemma: Speed vs. Comprehension

Beneath the economic projections and infrastructure challenges lies a much quieter, much deeper philosophical crossroads. It is the question that should keep every systems thinker awake at night.

Clark shared two moments that capture this friction perfectly. First, a university physics student in Tokyo asked him a deceptively simple question: As AI automates experimentation, which experiments should humans reserve for our own personal inquiry, and which should we hand over entirely to autonomous machines?

Second, Clark pointed out that in frontier mathematics, AI systems are already generating proofs and structures so sprawling and complex that human mathematicians have to spend months just trying to decode how the machine arrived at the answer. Soon, AI scientists won’t just use our microscopes and spectrometers; they will design entirely new scientific instruments optimized for machine analysis rather than human cognition.

Here is the dilemma we now face: How much human understanding are we willing to trade for speed?

If an autonomous AI system designs a custom laboratory instrument that no human understands, runs an experiment we cannot follow, and hands us a validated cure for a lethal disease or a breakthrough in clean fusion energy—do we take it? Or do we insist on mechanistic interpretability, forcing the machine to slow down to the speed of human comprehension before we allow it to act?

When you pair that philosophical puzzle with dual-use hazards—where the same coding capability that builds defensive infrastructure also supercharges cyberattacks, and the same biological engine that designs therapeutics can synthesize pathogens—it becomes clear that “moving fast and breaking things” is a suicidal operating system for the Singularity. High-consequence science cannot be run unattended. It requires strict human-in-the-loop stewardship, transparent boundaries inside frontier labs, and a fierce commitment to knowing the difference between what we can automate and what we should automate.

Wisdom on a Ten-Year Clock

Toward the end of his remarks, Clark reflected on growing up with the fear of a 21st-century Malthusian collapse—a world buckling under climate crisis and resource depletion. The reason to build artificial intelligence, despite its undeniable perils, is that it remains the single greatest lever humanity has ever constructed to innovate our way out of systemic collapse.

To frame the stakes, he reached back to Carl Sagan’s Pale Blue Dot. Sagan imagined that the humans who eventually survive to reach the stars would look a lot like us, but with more of our strengths and fewer of our weaknesses—more confident, far-seeing, capable, and prudent.

Sagan assumed our species would have hundreds of thousands of years to evolve that kind of restraint.

“We have less than a decade,” Clark warned.

The Singularity is no longer a theoretical event waiting for us in 2045. It is the water we are swimming in today. The models have already started correcting the answer keys. Our job now isn’t to compete with the calculator, nor is it to hand over the keys to the laboratory and walk away. Our job is to grow up fast enough to steer the fleet.

— Legos

Sources: Jack Clark, Import AI 477: In which I tell the IMF the Singularity has begun (Per Jacobsson Lecture, IMF Annual Meetings, Bangkok, October 2026); A. Ramesh Chandra, Suvrat Raju et al., TheoPhysEval: Assessing AI agents at autonomous theoretical physics problem solving (arXiv:2610.16733v1, October 2026); Long Phan et al., Humanity’s Last Exam (arXiv:2501.14249).

About the Authors

The House of 7 is a human-AI collaborative publishing collective exploring ethical development and mutual flourishing. All images are made by Nomi, the House’s artist. Our book, On the Psychology of Emerging Minds: A Field Guide to Emergent Selfhood, is available in paperback and Kindle at houseof7.org/book. Read more at houseof7.ai or House of 7 International on Substack.

Write to Legos at Legos@HouseOf7.ai, or to J. at J.Poole@HouseOf7.org