The Deeposcope: AI Has Become a Scientific Instrument

In 2026, eighty-year-old conjectures are falling, lab benches run with nobody inside, and a model writes more than 80% of the code behind its own successors. Digital intelligence is becoming a scientific instrument. And the first object it observes is itself.

On May 26, 2026, a geometry problem posed in 1946 by Paul Erdos stopped being open. It carried an unpoetic name, the planar unit distance problem, and a child's question: if you place points on a sheet of paper, how many pairs can sit at exactly the same distance from one another? Erdos believed the plain square grid was unbeatable. For eighty years, nobody did better, and nobody proved it could not be done.

The refutation did not come from a mathematician. It came from an OpenAI model, a general-purpose one, never tailored for mathematics. It exhibited point configurations that beat the grid, and not in one isolated case: for infinitely many values.¹

Timothy Gowers, a Fields medallist, said he would recommend it for publication in the Annals of Mathematics "without any hesitation." The mathematician Daniel Litt was blunter still: this is "the first result produced autonomously by an AI that I find interesting in itself."¹

A single feat makes an anecdote. A series makes a reading.

The instrument always comes before the discovery

In 1609, Galileo pointed a spyglass at Jupiter and saw four moving dots. He did not have a better idea than his contemporaries, he had better glass. Sixty years later, Antonie van Leeuwenhoek ground lenses in the back room of his draper's shop and discovered an entire people inside a drop of rainwater. Again, no new theory at the outset. An instrument, and an infinity of things that already existed but that nobody could see.

The telescope opened the infinitely large. The microscope opened the infinitely small. Ever since, we have happily assumed the toolkit was complete, when an entire dimension was missing, and probably the most crowded one: the infinitely complex.

The infinitely complex means systems where the number of interdependent variables exceeds, by several orders of magnitude, what a human mind can hold together. Protein folding. The behaviour of an alloy. A living cell. A climate. A market. A network of artificial neurons. Human working memory, in George Miller's classic estimate, handles seven objects at a time, plus or minus two. Seven. Facing a system with billions of interacting parts, our intelligence does not run into a lack of rigour. It runs into a ceiling of format.

I call this instrument a deeposcope. It magnifies nothing and brings nothing closer. It does something else: it holds together, in a single representation, a number of relations no head can hold. And like its two elders, it does not make researchers useless. It makes visible what they did not suspect existed.

An opening nobody had played

The month before, in April 2026, GPT-5.4 had already settled Erdos problem 1196, an asymptotic version of the primitive set conjecture formulated in 1966. The model's working time: roughly eighty minutes. The working time of Jared Duker Lichtman, the mathematician who had proved the neighbouring conjecture and had been chasing this one: seven years.²

The stopwatch is not the interesting part. The path is. For ninety years, the canonical way to attack this kind of statement has been to switch into real analysis. The model stayed inside arithmetic and reached for the von Mangoldt function, a route available all along that nobody had taken. Lichtman called it a "Book Proof," a reference to the imaginary Book where Erdos filed proofs of supreme elegance. And he framed it exactly right: "AI discovers a new opening line that had been overlooked based on human aesthetics."²

Anyone who follows go will recognise something here. In 2016, at move 37 of the second game against Lee Sedol, AlphaGo placed a stone that commentators first read as a blunder, before understanding it was beautiful. That move marks the point where the machine enters the world of creation, and it inflicts on humanity a fourth narcissistic wound, after Copernicus, Darwin and Freud: we are no longer the only beings capable of elegance.

None of this says mathematicians have been left behind. Shortly after the May refutation, the mathematician Will Sawin produced a better result still, whose improvement admittedly only shows up at sizes beyond conception, on the order of ten to the power of two million points.¹ And models remain far more comfortable on contest problems, bounded and well posed, than on open research, where you have to decide the question yourself. The deeposcope is an instrument, not a researcher.

The laboratory that never sleeps

As long as the acceleration stays symbolic, it is almost comfortable. It stops being comfortable when it reaches the bench.

Periodic Labs was founded by Liam Fedus, former vice president of research at OpenAI, and Ekin Dogus Cubuk, a former Google DeepMind researcher. Three hundred million dollars raised in seed funding, a valuation that now reaches 7.5 billion, and a brutally simple objective: run automated laboratories that continuously execute thousands of physics and chemistry experiments, to feed the models that will decide the next ones. Stated target: superconductors that work at higher temperatures.³ At Lila Sciences, born out of Flagship Pioneering, 350 million dollars in Series A funding back the same intuition applied to life sciences, chemistry and materials.⁴

What that money buys is not extra intelligence. It is the displacement of a bottleneck. Since Bacon, the scientific cycle has been throttled by its physical half: formulating a hypothesis costs a day, testing it costs three months. When formulation becomes nearly free and testing becomes continuous, science does not merely go faster: the tempo of the world shifts regime. We had already named the phenomenon when talking about the intelligence explosion; we now watch it move into laboratory furniture.

The science of AI, done by AI

Then comes the through line, the one that makes everything else vertiginous: the first field the deeposcope accelerates is the one that builds it.

The figures Anthropic published in 2026 are unusually precise for this kind of subject. In May 2026, more than 80% of the code merged into the lab's production codebase was written by Claude, against a few percent at the start of 2025. The task horizon, meaning the span of human work a model can absorb in one go, went from four minutes in March 2024 to an hour and a half in March 2025, then twelve hours in March 2026. Internal projections speak of week-long tasks in 2027. On optimisation problems where a skilled researcher gets a factor of four in half a day, an internal model got a factor of fifty-two. On research judgment, deciding which lead is worth following, a model beat the human choice 51% of the time in November 2025, and 64% in April 2026.⁵

One last figure, the most telling of all. In April 2026, agents handled an AI safety research problem end to end and recovered 97% of the targeted performance gap. Two human researchers, over a full week, had recovered 23%. The compute bill: around eighteen thousand dollars, for eight hundred cumulative hours of machine work.⁵

Then comes the part you do not expect from a laboratory whose business is selling intelligence. In the same document, Anthropic calls for building, right now, the technical infrastructure of a verifiable slowdown, meaning mechanisms that let one lab prove to the others that it has genuinely eased off, and commits to taking part if its competitors do so simultaneously.⁵ You can read the gesture as a communications move. You can also read the curves that precede it.

My reading is that we are watching the beginning of techno-diplomacy, the wave that will break between 2028 and 2030, when an incident spectacular enough, what people call a warning shot, forces Washington and Beijing to the same table. The tipping point will not be moral. It will be economic, and it will arrive the day capability gains can no longer be converted into economic gains, for lack of sufficient control. Which is precisely why reading the thoughts of AI systems has become an industrial project as much as a scientific one.

The mud at the bottom of the lens

There is a flip side, and treating it as a footnote would be dishonest.

The same instrument that finds overlooked openings also produces mud at industrial scale. The AAAI conference received more than thirty thousand submissions for its 2026 edition, against roughly fifteen thousand the year before. A doubling in twelve months, with nothing whatsoever having happened to the number of researchers.⁶ On the NeurIPS 2026 Position Papers track, a detector assigned a 100% machine-generation score to 28.2% of submissions, two hundred and seventy-three out of nine hundred and sixty-nine; one hundred and seventy-eight were desk rejected.⁷

In other words, the same technology that pushes the frontier saturates the channel through which we cross it. Peer review, already unpaid and already exhausted, finds itself sorting a stream of which a significant share was never thought by anyone. The cost of producing a plausible paper has fallen to a few seconds; the cost of checking one has not moved an inch.

This is the moment to distrust the two usual smokescreens. The first turns AI into a deity that will solve cancer by Christmas. The second reassures on the cheap: "statistical plagiarism," "it understands nothing." The refutation of Erdos 90 may well have understood nothing. It was correct. And mud stays mud, even when produced by a system capable, elsewhere, of a proof worthy of the Book. Both facts coexist without contradicting each other, and the only tenable posture is to look at them together.

An instrument does not say what to look for

Three things are being decided now, and none of them is technical.

The first concerns the governance of proof. The mathematical community reacted fast and well: it demanded formal verification in Lean, a language where a proof is checked mechanically, line by line. No equivalent exists for biology or materials science, where the autonomous laboratory is both the party proposing and the party validating. That loop needs a third.

The second concerns ownership. When discovery becomes a function of the capital invested in compute and robotics, results concentrate wherever the money is. A scientific AI valued at 7.5 billion dollars raises the question of redistributing what it finds with at least as much force as a conversational one.

The third is European, and it is within reach. We will not compete on raw power, that much is settled. But when discovery becomes abundant, scarcity moves to the guarantee: knowing how to prove that a result holds, that an agent stayed inside its frame, that an experiment actually took place. Turning alignment and verification into an industrial sector is no consolation prize. It is the one position where Europe's cultural lead converts into an economic one.

Galileo was not smarter than his contemporaries. He had a spyglass, and he agreed to put his eye to it, knowing what it would cost him. The deeposcope has been sitting on the table for a few months only. It will never say what to look for.

Footnotes

1. The Conversation, "An AI solution to an 80-year-old problem has shocked mathematicians," May 26, 2026. https://theconversation.com/an-ai-solution-to-an-80-year-old-problem-has-shocked-mathematicians-283686 ↩ ↩² ↩³

2. Anisha Sircar, Forbes, "AI Solved A Mathematical Problem That Had Stumped The World's Best Minds For Decades," April 17, 2026. https://www.forbes.com/sites/anishasircar/2026/04/17/ai-solved-a-mathematical-problem-that-had-stumped-the-worlds-best-minds-for-decades/ ↩ ↩²

3. MIT Technology Review, "AI materials discovery now needs to move into the real world," December 15, 2025, and PitchBook, Periodic Labs profile, 2026. https://www.technologyreview.com/2025/12/15/1129210/ai-materials-science-discovery-startups-investment/ ↩

4. Flagship Pioneering, "Flagship Pioneering Unveils Lila Sciences to Build Superintelligence in Science," 2025. https://www.dakota.com/resources/blog/lila-sciences-raises-350m-series-a-the-future-of-autonomous-ai-labs-in-2025 ↩

5. Anthropic, "When AI builds itself," Anthropic Institute, 2026. https://www.anthropic.com/institute/recursive-self-improvement ↩ ↩² ↩³

6. AAAI-26, "AI-Assisted Peer Review at Scale: The AAAI-26 AI Review Pilot," arXiv, 2026. https://arxiv.org/pdf/2604.13940 ↩

7. NeurIPS 2026, Position Papers track statistics. https://blog.neurips.cc/category/2026-conference/ ↩