The AI Spotlight Follow
AI Cracked a Million-Dollar Math Problem. Physics Is Next, but There Is One Job Machines Still Can't Do
Big Ideas

AI Cracked a Million-Dollar Math Problem. Physics Is Next, but There Is One Job Machines Still Can't Do

OpenAI's Navier-Stokes proof, a Gemini-powered cosmic string calculation and AI-designed gravitational wave detectors all point one way. Here is a step-by-step tour of what AI can already do in physics, where it still needs a human, and the missing skill researchers call scientific taste.

Advertisement

In September, OpenAI said an internal system had resolved one of the seven Millennium Prize Problems, a list of the hardest open questions in mathematics. Each one carries a $1 million prize. The system worked for about 88 hours, and its proof was checked by a computer program called Lean in another 17.

Mathematicians are still arguing about what it means. And a bigger question is hiding behind the argument: if AI is now breaking mathematics, will physics fall next?

The honest answer is more interesting than "yes" or "no." AI can already do many of the individual jobs a physicist does. What it can't do yet is the job that holds all the others together.

First, what OpenAI actually claimed

On September 8, 2026, OpenAI published a post titled "On the Navier–Stokes Millennium Prize Problem." The Navier-Stokes equations describe how fluids move: air over a wing, water in a pipe, smoke in a room. The famous question is whether smooth starting conditions can ever evolve into a "singularity," a point where something like fluid speed shoots to infinity in finite time.

Advertisement

OpenAI said a system of coordinating agents, on the order of 10,000 concurrent agents in the group that cracked it, found the answer in about 88 hours after launch. Formalizing and verifying the proof in Lean took another 17 hours using GPT-6 Astra. OpenAI also said it does not intend to claim the prize.

Now the important fine print, because it is a perfect example of how to read AI-and-science news:

  • The proof shows a blow-up for Navier-Stokes with an external smooth force applied. The official Clay Institute problem statement includes that variant as an allowed formulation (options "C" and "D"), so the result counts under the rules as written.
  • Many mathematicians care about the unforced version. Reporting in Scientific American says a later proof by three mathematicians showed that OpenAI's method cannot be extended to remove the force. Maths-outreach site Plus Magazine quotes one expert saying the problem "stated in its cleanest form has not been solved."
  • Lean confirmed the logic. But the paper runs 166 pages, and one mathematical physicist told Science News he doesn't think anyone has "completely verified the proof yet, certainly not on the human side."
  • The Clay Institute only awards the money once a peer-reviewed paper appears in a journal.
  • OpenAI says it learned during the project that a team involving an Anthropic employee and an NYU professor had separately resolved the forced Euler problem, and it acknowledged their priority on that one.

So a more accurate headline than "AI solved a Millennium Problem" is: AI produced a machine-checked proof that satisfies the official statement, and the human argument over whether it answers the question we really cared about is still going.

And note what it is not: it isn't a claim about real fluids. As one mathematician pointed out, the singularity only appears where the equations stop describing physical reality anyway, below the molecular length scale. That distinction between mathematics and physics is exactly where the rest of this story lives.

Is that physics, though?

Navier-Stokes is a math problem about equations that come from physics. A singularity of infinite density isn't something you'll ever measure. So many physicists shrugged.

But the line blurs quickly. Consider an example from earlier this year.

The cosmic string calculation

Cosmic strings are hypothetical, extremely thin structures that may have formed in the early universe. If they exist, wiggling loops of them would emit gravitational waves. Predicting the spectrum of that radiation boils down to evaluating a nasty integral over a sphere, one with singularities at certain points that make standard numerical integration unstable.

In a March 2026 paper from Google researchers, a system combining Gemini Deep Think with a systematic tree search and automated numerical checks found several new, exact analytic solutions. According to the paper, the agent identified a suite of six different methods, and the most elegant one expanded the integrand in a family of functions called Gegenbauer polynomials, whose weight naturally absorbs the troublesome singularities. The authors describe it as the system autonomously solving an open problem in theoretical physics.

This feels closer to physics than Navier-Stokes because the result predicts something measurable in principle: gravitational radiation from a cosmic string model. Pulsar timing arrays have renewed interest in exactly that kind of signal.

But look at what the AI actually did. Humans had to:

  1. Propose that cosmic strings exist at all.
  2. Work out what measurable consequence to look for.
  3. Translate the physics into that particular integral.
  4. Interpret the answer and weave it back into a picture of the universe.

The machine did step in the middle: the hard calculus. Which suggests a useful way to think about the whole question. Physics isn't one skill. It's a pipeline, and AI is entering it one stage at a time.

The pipeline: what AI can already do, stage by stage

Here is the tour, roughly in the order the technology arrived.

Stage 1: Pattern spotting and speed-ups (before deep learning)

Until about a decade ago, machine learning in physics meant boosted decision trees, support vector machines and shallow neural networks. They classified particle collisions and astronomical sources and learned complicated boundaries in data. Useful, but physicists still chose the representation: which quantities to measure, which variables mattered. The algorithm just found relationships among the variables it was handed.

Stage 2: Learning the representation itself

"Representation" means the mathematical encoding of a problem: which coordinates and variables describe a system, even what kind of equations to use. Neural networks loosened the requirement that physicists specify this in advance.

An early example is quantum mechanics. The state of a many-particle quantum system is described by a wave function, and the information needed grows catastrophically with the number of particles. In 2017, Giuseppe Carleo and Matthias Troyer showed that a neural network, a simple one with a single hidden layer called a restricted Boltzmann machine, could serve as a flexible representation of a many-body quantum state.

A few years later, FermiNet pushed the idea for electrons in atoms and molecules. Traditional methods get unwieldy when many strongly interacting electrons are involved. FermiNet baked the required spin statistics directly into a deep network and left the rest of the wave function flexible, so the network could feel its way to the lowest-energy state.

The recurring recipe in scientific machine learning: give the machine the physics you already trust, and let it learn the part you don't know how to represent.

That's powerful, and it contains a trap. The things you hard-wire are called inductive bias, information about the kind of answer you expect. In FermiNet's case, physicists supplied that electron wave functions are antisymmetric and obey the Schrödinger equation. Those are safe bets. But a great physicist's best trick is spotting when an assumption is wrong. Ptolemy preferred an Earth-centered cosmos. Copernicus saw through it, yet kept the bias for perfectly circular orbits, which Kepler later overturned. A truly general AI physicist would need to question even its foundations.

Stage 3: Rediscovering the laws of physics

How little can you tell a machine and still get physics back? The landmark answer is AI Feynman, developed by Silviu-Marian Udrescu and Max Tegmark. It is a hybrid: a neural network helps expose hidden structure, such as whether a relationship splits into simple pieces or has symmetries, dimensional analysis checks units, and symbolic regression searches through explicit math expressions until it finds a compact equation.

On a benchmark of 100 equations drawn from the Feynman Lectures on Physics, it recovered all 100. A popular competing tool, Eureqa, solved 71.

The hybrid design is itself a lesson: you probably can't duplicate everything a physicist does with one architecture, but several complementary methods working together might.

AI Feynman still needed hand-holding. People chose the variables, provided clean numbers and said roughly what kind of equation counted as an answer. Newer systems relax that. One approach starts from raw video, learns a small set of hidden coordinates in which the motion looks simple and then runs symbolic regression on them. It finds the representation and the law at the same time. It isn't assumption-free: it effectively applies Occam's razor, preferring low-dimensional variables and simple equations. A more recent system, Pixel2Phys, uses multimodal foundation-model agents on simulated and real scientific video: different parts of the system propose variables, derive equations, test them against the observed motion, and revise them when they fail.

Here is the peculiar trade-off. The systems given the fewest assumptions mostly rediscover laws we already know. That validates the approach. It isn't new physics. To do useful new physics today, you tend to build in more scaffolding and more bias, which means progress in directions we already pointed.

Stage 4: Designing the experiments

Physics isn't only theory. Many quantum breakthroughs came from bouncing photons around a tabletop, and the number of possible arrangements of lasers, mirrors and beam splitters is beyond human exploration.

In 2016 a system called Melvin, by Mario Krenn and colleagues, treated quantum optics experiments as a combinatorial search problem, learning and reusing useful building blocks from earlier searches. It found unusual configurations human researchers had missed.

The much bigger example is Urania, from Krenn, Yehonathan Drori and Rana Adhikari, published in Physical Review X in 2025. It searches the space of gravitational wave detector layouts using gradient-based optimization. Starting from LIGO Voyager as a baseline, it found dozens of unconventional designs that beat the human-optimized version, 50 of which were published in a public "detector zoo." The authors report that the best could increase the observable volume of the universe by up to about 50-fold, and one post-merger design by roughly 69-fold in its target frequency band. Two caveats from the project itself: none of these designs has been built, and constraints like manufacturing haven't been fully accounted for.

Stage 5: Transformers, reasoning and agents

Then came transformers, the architecture behind modern language models. Their key feature is attention: the ability to connect an equation to a definition many paragraphs earlier, or a line of code to the variable it uses. A bigger context window means the model can hold more of a paper, a derivation or a codebase in view at once. For science, where problems depend on long-range connections between elements, that matters a lot.

Pre-train a transformer on huge datasets, then post-train it with reinforcement learning, and it can learn control policies. One earlier example: a neural network learned in simulation to control the magnetic coils of the TCV tokamak, then transferred that policy to the real fusion device and held a variety of plasma shapes.

But reinforcement learning works best when a task is repeated actions with a clear reward, and research is messier. Humans are good at long-horizon planning and synthesis, and we are contemplative: we can inspect our own thought. The fix is inference-time compute: train models to think carefully with more internal iteration, or wrap them in a scaffold that iterates over their outputs. The result is a reasoning model.

Put that into an agentic system with access to code, compilers, simulators, data tools and the scientific literature, and you get the cosmic string result: a model proposing approaches, a tree search branching into strategies, and numerical tools checking each step. Some systems are even being connected to real instruments. Early versions have begun carrying out pieces of actual experiments at synchrotron facilities, such as X-ray crystal alignment.

So can AI already do physics?

To a large extent, yes, in pieces. Today there are systems that help infer unknown laws from observations, invent experimental apparatus, recognize reusable structures across many solutions, choose mathematical strategies, operate scientific instruments and contribute to active theoretical research.

Several teams are building end-to-end "AI scientists," and a few are building explicit AI physicists. The open question is what's left for humans.

The one thing that's still missing

A 2024 paper by Ruairidh Battleday and Samuel Gershman, now published in Philosophical Transactions of the Royal Society A, splits AI for science into an easy problem and a hard problem.

  • The easy problem (which they concede is itself hard) is performing the steps of science: optimizing a pre-specified objective, answering the questions posed, predicting protein structures, discovering antibiotics.
  • The hard problem is formulating the right question in the first place: deciding which phenomena matter and what constraints a theory should satisfy.

A 2026 study called AstroInsight tested this in astronomy. Agentic language-model systems read the literature and generated research ideas that human experts then judged. Some ideas were genuinely novel and promising, but they were mostly useful extensions and recombinations of existing approaches. The study found no evidence of directions that would lead to true breakthroughs.

That's combinatorial novelty: absorbing a field and reconfiguring the parameters of past work to find what hasn't been done. Real breakthroughs need frameworks to be broken. Einstein asked what if space and time are mutable and gravity is equivalent to acceleration. Heisenberg asked what if we describe only the observables of the atom. A machine could in principle riff off past insights like those, but current AI looks too framework-locked to ask a question that no one has asked.

Researchers sometimes call the missing ingredient scientific taste: an instinct for which questions are worth asking. It blends deep knowledge with curiosity, imagination, abstraction, sensitivity to anomalies, skepticism and even aesthetics. AI has pieces of each, but doesn't yet integrate them the way a good scientist does.

Understanding is the last mile

There's a second, quieter point. Suppose an AI found the master equation, the theory of everything. If no one understands what it means, who benefits? Physics isn't just a list of correct formulas. It's an integrated worldview humans have built over millennia, and new results have to be woven back into it. For now, that integration is the job humans do, and some argue it may always be needed.

The sober forecast, which is one scientist's view rather than a settled fact: AI is unlikely to solve physics by itself. Physics will change, for better and worse, and AI is currently most useful when directed by human physicists in service of human understanding.

Why this matters

  • Credit and verification are about to get complicated. The Navier-Stokes story shows a 166-page proof checked by software before most humans had read it. Journals and prize committees will need new norms.
  • Careers are shifting. If calculations, equation-finding and experiment design are increasingly automated, the premium moves to asking good questions, judging results and explaining them.
  • Read the claim, not the headline. "Solved the Millennium Problem" and "solved a forced variant that satisfies the official statement" are very different sentences.

What this means for you

  • If you're a student or researcher, learn to direct and audit AI tools: check outputs, understand the inductive biases built in and be able to say why an answer is right.
  • If you follow AI science news, ask three things: what exactly was proved or found, who checked it and what assumption was built in?
  • If you build AI tools, remember the AI Feynman lesson: hybrids of neural networks, symbolic search and physics constraints often beat any single model.

FAQ

Did AI really solve the Navier-Stokes Millennium Problem? OpenAI says its system produced a machine-verified proof that Navier-Stokes solutions with an external force can blow up in finite time, which fits the official Clay statement. Experts dispute whether that answers the version most mathematicians care about, and human verification of the long proof is ongoing.

What did AI do with cosmic strings? A Google research system using Gemini Deep Think plus tree search found exact analytic solutions to a hard integral describing gravitational radiation from cosmic string loops. Humans supplied the physical setup and interpretation.

Can AI discover new laws of physics? So far, mostly it rediscovers known laws from data, as AI Feynman did with all 100 test equations. Truly new physics has needed more human-provided structure.

What is "scientific taste"? A researcher's instinct for which questions are worth pursuing. Current AI is good at answering questions and recombining known ideas but weak at generating paradigm-shifting new questions.

Will AI replace physicists? Nobody knows. The evidence so far says AI is taking over many individual steps while humans still do question-setting and integration.

The bottom line

AI hasn't solved physics, and it hasn't solved the Navier-Stokes problem in the way the headline suggested. But it is quietly taking over the pipeline one stage at a time: the calculus, the equation-hunting, the experiment design, even some of the lab work. What's left is the part that decides what any of it is for. The first AI that asks a question no human has thought to ask will be the real milestone, and so far, nobody has shown one.

Sources

AI Is Doing More of Your Thinking. The Research on What That Costs Your Brain Is Getting Hard to IgnoreUp next · 11 min readAI Is Doing More of Your Thinking. The Research on What That Costs Your Brain Is Getting Hard to IgnoreBrain scans of students writing with ChatGPT, a flood of 'workslop' at the office, a sneaker company that became an AI stock overnight: the evidence that we're outsourcing judgment, not just typing, keeps piling up. Here's what the research actually shows, why groupthink makes it worse, and a practical plan to keep your mind your own.Keep reading →