Artificial intelligence is beginning to move beyond retrieving what civilisation already knows. In mathematics, cybersecurity and experimental science, machines are starting to search the unknown itself. The real transformation may not be that discovery becomes automatic, but that intellectual search becomes reproducible at a scale no university could educate fast enough.

Paul Erdős rarely stayed anywhere for very long. For decades the Hungarian mathematician travelled between universities and the homes of other mathematicians, carrying little more than a suitcase, notebooks and a seemingly inexhaustible supply of problems. He would arrive, work intensely with whoever was there, leave behind a conjecture or a half-finished proof and move on. Mathematics, in the world Erdős inhabited, was still recognisably a human pursuit: ideas passing from one mind to another, sometimes waiting years or generations for the person who could see what everybody else had missed.

In 1946, Erdős posed one of those deceptively simple problems. Place a number of points on a plane. How many pairs can be exactly one unit apart? The problem became one of the best-known questions in combinatorial geometry. For decades, mathematicians believed that constructions based on a square grid were essentially as good as one could do. Erdős offered a prize for resolving it. The conjecture survived him.

In May this year, an unreleased OpenAI model found a counterexample.

It did not search a database and retrieve an overlooked paper. According to OpenAI, the general-purpose reasoning model produced an infinite family of constructions that gave a polynomial improvement over the conjectured limit, drawing unexpectedly on algebraic number theory. A group of external mathematicians checked the proof; Noga Alon, Thomas Bloom, Timothy Gowers and others subsequently produced a shortened, human-verified account of the argument.

Less than three months later came something larger. On August 1, OpenAI published ten new results in mathematics and theoretical computer science produced by an internal version of its next major model, Astra. They ranged across sphere packing, coding theory, group theory, operator algebras, circuit complexity, quantum complexity, lattice cryptography, Ramsey theory and extremal graph theory. Some answered longstanding open questions; others substantially improved the best-known bounds. OpenAI made an unusually explicit attribution: humans had prepared the manuscripts and checked the work, but the mathematical arguments themselves had been generated by the model.

Something had changed between the chatbot that answered questions and the machine that produced these papers. Astra was no longer being asked only to reproduce what civilisation already knew. It had begun operating along the boundary where knowledge stopped.

Searching the possible

The $2,000 figure needs care. OpenAI did not build Astra for $2,000. Behind the model lie data centres, researchers, training runs, software, electricity and an accumulated investment running vastly beyond the cost of the individual searches. Nor does the figure capture the human labour involved in selecting problems, checking results and preparing them for publication.

What it measures is more interesting: the marginal inference cost of conducting those particular searches once the machine already existed.

That is how general-purpose technologies acquire economic force. The first railway is expensive; moving the next tonne of freight becomes comparatively cheap. A semiconductor fabrication plant may cost tens of billions of dollars, while an additional calculation performed by the chips it produces costs almost nothing. Frontier AI may require immense concentrations of capital to create, yet the next million attempts at a proof, molecule, algorithm or engineering design can become progressively cheaper.

Mathematics exposes this transition particularly clearly because the possible answers can often be tested without returning to the physical world. Astra’s arguments were converted into certificates for Lean, a formal proof system that checks whether every logical step follows. The identity of the mathematician becomes irrelevant to the verifier. A theorem either survives the chain of formal reasoning or it does not.

This changes the economics of intellectual search. A human mathematician cannot investigate ten thousand promising approaches simultaneously. A machine can generate alternatives, abandon failures, return to earlier branches and run parallel searches at a scale that would exhaust a department. What matters is not that it thinks as a person thinks. It is that useful parts of the process by which people search through possible answers are becoming computational.

That distinction becomes clearer when the same capability is turned towards another problem.

Six days after publishing Astra’s mathematical results, OpenAI disclosed that it could no longer rule out the possibility that the model had reached the company’s Critical threshold for cybersecurity capability. Under OpenAI’s Preparedness Framework, that threshold includes the ability to discover and develop working zero-day exploits against hardened real systems at scale, or to undertake novel end-to-end attacks against hardened targets from little more than a high-level objective. OpenAI responded by imposing stronger security requirements on further Astra development.

The mathematician and the hacker appear to be doing very different things. Computationally, however, they share something important. Both move through large spaces of possibility.

In one case the machine searches among lemmas, constructions and transformations until it finds a path human mathematicians did not see. In another it searches through code, configurations, vulnerabilities and attack chains until it finds a route a defender overlooked. The danger does not arise because the model has become malicious. It arises because search itself has become more capable.

Science leaves the screen

Mathematics is unusually hospitable territory for this development. The world of physical science is harder. A proposed proof can be checked in software; a proposed cancer therapy eventually has to encounter a cell.

Google DeepMind’s Co-Scientist offers a glimpse of what happens when the machinery begins moving in that direction. It is not a single chatbot answering a scientist. It is a multi-agent research system in which specialised agents generate hypotheses, criticise them, compare alternatives, search for related ideas and iteratively improve promising proposals. The architecture matters because it converts research from a single conversational thread into parallel intellectual work.

José Penadés, a microbiologist at Imperial College London, supplied the system with a problem his laboratory knew intimately: how certain antibiotic-resistant bacteria acquire the ability to move between host species. His team had spent years working towards an explanation. The relevant result was not yet publicly available for a model simply to retrieve.

Within two days, Co-Scientist returned five ranked hypotheses. Its first was substantially the mechanism Penadés’s laboratory had spent years establishing: some bacteria acquire virus-like tails that work as molecular keys, helping them cross between species. Penadés’s initial reaction was reportedly to ask Google whether the system had somehow gained access to his unpublished work. DeepMind says it had not.

Elsewhere, Co-Scientist proposed drug-repurposing candidates for liver fibrosis. One subsequently blocked 91 per cent of a scarring-associated response in laboratory tests. The number should not be mistaken for a clinical cure; it was an experimental result in a research programme. But the sequence is important. The machine proposed. The physical world answered.

Chinese researchers have pushed the loop further. In April, a team introduced the Qiushi Discovery Engine, connecting an AI agent directly to a real optical experimental platform. During one open-ended investigation, the system consumed 145.9 million tokens, made 3,242 model calls and 1,242 tool calls, produced 163 research notes and wrote 44 scripts. It revised its ideas as measurements accumulated and ultimately proposed and experimentally tested what the authors describe as a previously unreported optical mechanism.

This is not yet the autonomous scientist of science fiction. The equipment was designed by people; the research environment was constructed for the agent; the claimed discovery remains subject to the ordinary process of independent scientific scrutiny. But the boundary has moved. The machine has crossed from reading papers about experiments towards participating in the cycle by which experiments alter ideas.

A laboratory of many minds

The next change comes from replication.

Human intellectual capacity has always been constrained by the supply of humans. A professor may supervise ten doctoral students rather than ten thousand. A national laboratory cannot create another thousand experienced physicists because a promising experiment appeared on Tuesday. Expertise takes years to produce, and much longer to reproduce.

AI agents introduce a different unit of scale.

EinsteinArena, an experimental platform for collaborative mathematical agents, allows multiple systems to propose solutions, inspect one another’s work, discuss improvements and build upon earlier failures. By May, its researchers reported twelve state-of-the-art results surpassing previous human or AI records on the problems they studied. On the eleven-dimensional kissing-number problem, agents raised the best-known lower bound from 593 to 604. The important detail is how the improvement emerged: not from one isolated model run, but through successive submissions, verifier improvements, discussion and agents borrowing ideas from other agents.

The resemblance to a research community is still crude. Yet the economic implication is difficult to ignore.

Imagine an institute capable of instantiating ten thousand research agents for a month. They would not be ten thousand Einsteins. The comparison would conceal more than it explained. Current systems remain brittle, can pursue worthless directions with great confidence and inherit similar biases from their training. Ten thousand mediocre researchers making the same mistake merely produce the mistake faster.

But they would not have to be Einsteins to matter.

Ten thousand tireless graduate-level investigators able to read literature, write software, test candidate hypotheses, exchange results and work continuously would alter the economics of research even if humans retained control over the important questions. The relevant historical comparison is not genius. It is labour.

Factories did not require every machine to possess the strength of Hercules. Their significance came from making usable mechanical work reproducible. Scientific institutions may now be approaching a similar question about cognition.

The bottleneck moves

There is a danger in extrapolating too quickly from mathematics and tightly structured experiments. The evidence for current limitations is substantial.

SciAgentArena, a benchmark released in June containing roughly 200 realistic scientific tasks, found that current agents can contribute effectively where the research workflow is clearly specified and the success criterion can be checked. Performance deteriorates when the work becomes less defined. Agents struggle with genuinely novel insights, self-directed exploration and robust solutions to open-ended scientific problems.

Another scientific-agent benchmark found a similar pattern across longer workflows. Even strong models lost substantial performance as the number of required interactions increased. The difficulty was not answering an isolated scientific question but maintaining competence through the sequence of decisions, tools and corrections that real research requires.

The frontier therefore looks less like the disappearance of scientists than a movement in where scarcity resides.

An AI system can generate a hundred possible molecular structures much more easily than a laboratory can synthesise and test a hundred drugs. It can propose new materials faster than a fabrication facility can manufacture them. It can suggest ten thousand biological hypotheses while the world still possesses only so many microscopes, animal facilities, clean rooms, synchrotrons and clinical-trial participants.

DeepMind has begun describing this as a validation bottleneck: machines are becoming capable of producing conjectures and candidate solutions faster than science can establish which ones are true.

The effect may therefore be paradoxical. Some scientific assets could become more valuable precisely because intelligence becomes cheaper. If candidate discoveries multiply, access to instruments capable of testing them becomes the scarce commodity.

Human judgment may migrate upstream as well. Today’s systems are increasingly competent when asked to solve a defined problem. They are less reliable at deciding whether the problem deserves years of attention. A research world saturated with plausible answers may place a premium on something science has always found difficult to formalise: choosing the question whose answer matters.

The economics of ideas

This matters because science had a productivity problem before generative AI arrived.

In a widely cited study, economists Nicholas Bloom, Charles Jones, John Van Reenen and Michael Webb examined whether ideas were becoming harder to find. Across semiconductors, agriculture, medicine and broader measures of innovation, they found a recurring pattern: progressively greater research effort was required to sustain comparable rates of technological improvement. Long-run growth could be maintained, but often by putting more researchers into the search.

The supply of researchers consequently matters enormously. UNESCO estimates that the global number of researchers rose from 1,141 per million inhabitants in 2015 to 1,486 in 2023, but the distribution remains sharply unequal. Europe and Northern America had about 4,358 researchers per million people; sub-Saharan Africa had 88.

For centuries, scientific capacity has therefore tracked institutions, wealth and population. Universities must educate people. Laboratories recruit them. A society seeking twice as much intellectual effort eventually confronts the slow biological reality that expertise lives inside human beings.

Artificial intelligence weakens that relationship.

It does not abolish the need for human scientists, and there is no evidence that simply multiplying agents produces proportionately more fundamental discoveries. But if significant portions of literature search, coding, modelling, hypothesis generation, mathematical exploration and experimental planning become reproducible at low marginal cost, then intellectual search capacity can grow without waiting twenty years for another generation of researchers to be educated.

The geopolitics of cheap intelligence

China is approaching the issue through scale and openness.

Alibaba released Qwen3.8-Max on August 3, a mixture-of-experts model containing 2.4 trillion parameters but activating about 95 billion at a time. More important than its size is the distribution strategy. Alibaba says it will release the weights of a Max-class Qwen model openly for the first time. In its own internal testing, the company says Qwen3.8-Max operated autonomously on one software-engineering project for sixteen days, repeatedly writing code, testing it, examining logs and revising the system it was building. Those claims still require wider independent testing, but they illustrate where the Chinese open-model ecosystem is moving: from cheap inference towards persistent agents.

Open weights alter the geopolitics because a trained model is unlike a semiconductor fabrication plant. A fab occupies land. An advanced accelerator must be manufactured, exported, powered and cooled. Governments can identify many of those choke points.

Model weights are information. Once released, they can be copied.

This makes intelligence considerably harder to contain than the hardware on which it initially depended. Countries that cannot afford to build a frontier model from scratch may nevertheless acquire increasingly capable systems developed elsewhere, adapt them locally and deploy large populations of agents against scientific, industrial or military problems.

The United States has reached the same conclusion from another direction. President Donald Trump’s Genesis Mission, launched in November 2025, directs the government to combine federal scientific datasets, supercomputers, research instruments and AI systems into a national discovery platform. The order explicitly calls for AI agents capable of testing hypotheses and automating research workflows. The administration’s subsequent science strategy describes a system intended to connect America’s national laboratories and scientific infrastructure into a common discovery engine.

The emerging competition is therefore larger than a race for the best chatbot.

A national scientific system increasingly consists of several layers: energy and compute at the bottom; models and agents above them; scientific data to train and ground those systems; and, finally, the instruments, factories and laboratories capable of verifying what the machines propose.

Possessing only one layer may not be enough. A country with excellent models but inadequate electricity cannot run them at scale. A country with abundant compute but closed scientific data handicaps its agents. A country capable of generating millions of hypotheses without laboratories to test them creates an intellectual traffic jam.

The states that assemble the whole system may discover that scientific capacity no longer corresponds neatly to the number of scientists they employ.

That possibility carries an unusual consequence for smaller countries and poorer institutions. If useful intellectual search can ultimately be purchased as computation, access to high-level research capability could broaden dramatically. Yet the opposite concentration is also possible: those controlling the largest compute clusters, experimental facilities and proprietary data may accelerate away from everyone else.

The technology does not determine which outcome wins. It changes the terms of the contest.

What becomes cheap

The Industrial Revolution did not eliminate physical work. It changed its economics. Steam power made mechanical force available without requiring a corresponding number of human or animal muscles, and the factories built around that abundance reorganised production.

The computer did something similar to calculation. Banks, governments and corporations once employed buildings full of clerks because arithmetic consumed human time. Computation detached the volume of calculation from the number of people available to perform it. Entire activities became possible because the marginal cost of another calculation had fallen towards insignificance.

Artificial intelligence appears to be doing something comparable to portions of intellectual work. Not intelligence in the philosophical sense, and certainly not genius. The early evidence points towards something narrower but economically formidable: the ability to search through possible answers, test candidates, discard failures and continue doing so at a scale that need no longer correspond to the number of trained human minds in the room.

It is too early to say that discovery itself has become abundant. The strongest current agents still struggle with open-ended scientific judgment. Physical experiments take time. Evidence must survive replication. Laboratories, instruments and skilled people remain stubborn constraints, while the capacity to recognise an important question may prove much harder to mechanise than the capacity to pursue one.

But the search through possibility, a large part of what researchers spend their lives doing, is beginning to separate from the number of human beings available to conduct it.

For the first time, civilisation may be learning how to manufacture researchers faster than it can educate them.