A proposed solution to a landmark mathematics problem and a map of nine billion possible DNA changes suggest AI is expanding the frontiers of research. The harder task may be establishing which of its answers deserve to become accepted knowledge.
For roughly 90 years, mathematicians have struggled with a question about the equations governing fluid motion. OpenAI says an artificial intelligence system found an answer in less than four days.
The claim, announced on September 8, concerns the Navier–Stokes existence and smoothness problem, one of mathematics’ most celebrated unresolved challenges. OpenAI has released a proposed proof and a computer-checkable version. Whether the result withstands independent scrutiny will determine its place in mathematical history. OpenAI’s announcement
On the same day, Google DeepMind announced AlphaGenome Atlas, a database predicting the molecular effects of approximately nine billion possible single-letter changes in human DNA. Its ambition is to help researchers identify consequential mutations in a search space too vast to investigate exhaustively in a laboratory. DeepMind’s announcement
One announcement offers a proposed mathematical answer; the other, an immense collection of biological predictions. Together, they sharpen a question that reaches beyond either discipline: what happens when machines can generate potential discoveries faster than scientists can check them?
Four days against a problem of generations
The Navier–Stokes equations describe the movement of liquids and gases, underpinning work from aircraft design to weather forecasting. Their usefulness is beyond dispute. Their ultimate mathematical behaviour has remained uncertain.
The question is whether a smooth, three-dimensional fluid flow must remain well behaved, or whether it can develop a singularity — a point at which quantities such as velocity become unbounded within a finite time.
OpenAI says its system constructed an example in which a smooth external force acts on a fluid initially at rest, producing an increasingly narrow and fast-moving vortex. Velocity grows without bound while total energy remains finite.
That is a claim about the behaviour of a mathematical model, not a prediction that ordinary water can suddenly move infinitely fast. The company says the construction satisfies the conditions of the forced versions of the Millennium Prize problem. OpenAI’s account of the result
The method is almost as striking as the claim. According to OpenAI, the group that produced the result involved about 10,000 AI agents. They explored competing approaches, with useful intermediate findings consolidated and shared across groups.
The agents reached their proposed solution after approximately 88 hours. Formalisation and verification using the Lean proof system took another 17.
Those figures describe a concentrated computational effort built on generations of human mathematics. They do not establish that a machine could recreate the discipline from scratch. But if the proof holds, the system has assembled an argument that was absent from the knowledge on which it drew.
The announcement is only the beginning
A corporate announcement cannot confer mathematical acceptance.
Computer verification is valuable because it can check whether a formal argument follows from specified assumptions. Researchers must still establish that those assumptions and definitions correctly capture the problem being claimed as solved.
The distinction matters particularly here. A counterexample involving an external force should not be read as settling every question about fluid motion, or as delivering an immediate practical solution to turbulence.
The strongest defensible claim is also compelling enough: OpenAI has presented an AI-generated proof of a major result and made it available for examination. Its significance now depends on that examination.
A map of where biology should look
DeepMind’s Atlas addresses a different obstacle: the overwhelming number of possible genetic changes.
The human genome contains roughly three billion DNA letters, each of which can be replaced by one of three alternatives. AlphaGenome Atlas supplies predictions across those approximately nine billion substitutions.
The resource is especially relevant to the roughly 98 per cent of the genome that does not directly encode proteins. Some of this DNA regulates when and where genes operate, and how their instructions are processed. Changes there can have substantial effects without directly altering a protein’s coding sequence.
The Atlas gives researchers a way to rank variants and investigate the molecular processes they might disrupt. It is a map for choosing experiments, rather than a catalogue of experimentally established facts.
An early example illustrates the distinction. According to DeepMind, researchers investigating rare disease used its scoring system to identify a variant affecting DNM1, a gene associated with severe neurological disease. AlphaGenome predicted that the variant would disrupt RNA splicing — the processing of genetic instructions — and produce an abnormal protein extension. Experimental work validated the predicted effect. DeepMind’s account of the research
The value lay in a testable explanation. The prediction became useful evidence when researchers took it into the laboratory.
Verification becomes the scarce resource
The broader implication is a change in the economics of research.
Scientific progress has long been limited by the number of promising ideas that people have the time and resources to pursue. AI could make parts of that search much cheaper and faster, while leaving the cost of establishing the truth stubbornly high.
Mathematics has an advantage: formal proof systems can automate substantial parts of logical checking. Biology requires experiments, replication and, where medical applications are concerned, clinical evidence. Generating another million plausible mechanisms does not create another million laboratories.
That imbalance could reshape research priorities. Scientists will need better ways to select which predictions deserve scarce experimental resources. Universities may face a growing disadvantage against companies able to finance enormous computational searches. Journals will have to distinguish productive machine-generated work from a rising volume of plausible but unsubstantiated claims.
Neither announcement establishes that scientists are becoming dispensable. Both suggest their judgement could become more consequential.
The promise is considerable: unexplained diseases, elusive mathematical constructions and useful biological mechanisms may become easier to find. But discovery earns its authority through scrutiny.
AI may accelerate the search for answers. Science will still have to establish which ones are true.