
AI Is Breaking Math. Will It Solve Physics Next?
Season 12 Episode 3 | 22m 29sVideo has Closed Captions
Could the next Einstein be a AI, or is there something fundamentally human about physics?
AI is solving math that has stumped humanity for decades. But can it do more than solve equations? From discovering hidden laws of nature to designing entirely new experiments, AI is beginning to take on the work of theoretical physicists. Could the next Einstein be a machine, or is there something fundamentally human about making the discoveries that change our understanding of the universe?
Problems playing video? | Closed Captioning Feedback
Problems playing video? | Closed Captioning Feedback

AI Is Breaking Math. Will It Solve Physics Next?
Season 12 Episode 3 | 22m 29sVideo has Closed Captions
AI is solving math that has stumped humanity for decades. But can it do more than solve equations? From discovering hidden laws of nature to designing entirely new experiments, AI is beginning to take on the work of theoretical physicists. Could the next Einstein be a machine, or is there something fundamentally human about making the discoveries that change our understanding of the universe?
Problems playing video? | Closed Captioning Feedback
Where to Watch PBS Space Time
PBS Space Time is available to stream on pbs.org and the PBS app.
Providing Support for PBS.org
Learn Moreabout PBS online sponsorshipThe AI revolution is changing many fields, both for good and bad.
Maths is in the midst of an AI driven existential shift.
Will physics follow fast on its heels?
When will we get our first true AI physicist?
And will it solve physics?
In September 2026, OpenAI announced that an internal AI system had produced a proof resolving one of the.. in mathematics.
The Navier-Stokes existence and smoothness.. The new Navier-Stokes equations describe the motion of fluids, of flowing air, of swirling water, and are an essential staple of fluid mechanics.
But fluids are chaotic and rarely permit simple solutions to their equations.
And worse, the Navier-Stokes equations hint at the possibility of a fluid swirling itself into a vortex of infinite density.
It's hard stuff, maths wise, and the mathematical proofs that these Navier-Stokes equations always permit globally smooth solutions, and that they suggest evolution into singularities in finite time, were designated as one of the seven Millennium Prizes in mathematics, and in September this year, OpenAI claimed that a team of their AI agents had proved the singularity aspect of the prize problem.
Now the claim still has to be verified, and there's also the claim that the AI system may have peaked at work by a couple of actual human mathematicians.
Regardless, it's a remarkable result.
This and other recent major breakthroughs in mathematics have mathematicians wondering how their own work is going to shift going forward.
But Navier-Stokes is physics, so should physicists be excited and or worried?
Also know this recent work is arguably not physics.
The Navier-Stokes singularity is definitely not physical, because those equations break down when we approach the scale of molecules and atoms.
But the distinction between doing maths and doing physics blurs very quickly in other situations.
Here's an example.
Earlier this year, an AI assisted system solved an open problem involving cosmic strings, hypothetical structures that may have formed in the early universe.
Calculating the gravitational waves produced by oscillating cosmic strings requires evaluating a difficult integral containing mathematical singularities.
A system combining Google's Gemini deep think with tree search and automated numerical checks found several new .. That paper claimed to have solved an open problem in real theoretical physics.
And this does sound more like physics than the Navier-Stokes thing, because the calculation predicts something physical that, in principle is testable.
The gravitational radiation produced by the cosmic string model, but the AI’s actual contribution, again, still looks a lot like doing maths and physics feels like it should involve a few additional steps in the cosmic string result.
Actual human physicists had to come up with the hypothesis, including the very existence of cosmic strings, to figure out what a measurable consequence might be like the gravitational wave spectrum, and then translate the problem into an integral, which could then be solved by the AI system.
But perhaps most importantly of all, physicists then needed to interpret the results of calculations and weave them back into an integrated understanding of.. But different developments in machine learning are beginning to reach into different parts of that broader scientific process.
And indeed, today we have systems that are hoping to infer unknown laws from observations, invent experimental .. recognize reasonable structures across many solutions, choose mathematical strategies, and even operate scientific instruments and contribute to active theoretical research.
AI is doing things that we once thought purely the province of the human physicist.
So to some extent, AI can already do physics.
But how much more of doing physics will we ultimately hand over to the machine?
Will we eventually or imminently have something that we can reasonably call an AI physicist?
Or will I remain forever fundamentally a tool with human physicists holding the reins?
To get at this, we're going to follow two developmental threads together what the AI is technically capable of and what scientific choice that capability allows us to hand over.
Up until around ten years ago, before the era of deep learn.. the most impressive uses of machine learning in physics mostly involved recognizing patterns in enormous data sets or speeding up calculations.
Boosted decision trees, support vector machines and shallow neural networks did things like classify particle collisions and astronomical sources that could also learn complicated boundaries in data that could point to hidden and unexpected relationships.
But physicists still chose which parameters should be used to represent the system position momentum of a particle, morphology of a cosmic ratio, or a galaxy, etc.
but given such a starting representation of the system, these algorithms could infer somewhat complex relationships between supply variables, which is like figuring out a formula, even if the framework was pretty restricted.
But representation learning marked an important transition in the boundary of what machines could do.
By representation, I mean the mathematical encoding of the problem.
What physical quantities describe a system state which variables and coordinates, and even the broad kind of equations or mathematical .. Neural networks loosened this requirement because they could learn usefu.. rather than forcing the physicist to specify the detailed mathematical form in advance.
Here's an early example.
In quantum mechanics, the state of a many particle quantum system is encoded in its wave function, and the amount of information needed to describe that state grows catastrophically with the number of particles.
In 2017, Giuseppe Carleo and Matthias Troyer showed that a neural network could itself serve as a flexible mathematical representation of a many body quantum state.
The network was a type called a restricted Boltzmann machine, and a simple neural network architecture with a single hidden layer that could effectively implement the complex wave function without requiring its mathematical form to be specified by the physicists.
A few years later, Fermi net pushed this idea further for el.. and molecules.
Now, traditional methods quickly become unw.. with many strongly interacting electrons.
In particular, because it's hard to keep track of the spin interactions of all electrons.
Fermi net baked the required spin statistics directly into a deep neural network, while leaving the rest of the wavefunction flexible enough to learn the complicated correlations.
The network could then feel its way to the lowest energy or ground state solut.. So the physicists hard wired one piece of physics that we know must be true.
While giving the network a lot of freedom to discover the rest of the wavefunction for itself, this illustrates a recurring strategy in scientific machine learning.
Give the machine the physics that you already trust, and let it learn the part that you don't know how to represent efficiently.
That means baking your assumptions into the learner, fixing what you think you already know about the physics.
But what if you're wrong about those assumptions?
Now the technical term is inductive bias.
Information about the kind of answer you expect supplied in advance.
Sometimes you are aware of the bias, but sometimes you aren't.
In the case of firming it, human physicists supplied the fact that the electrons have antisymmetric wave functions and that they obey the Schrodinger equation.
And I'd know that electrons even exist, and wh.. That all seems reasonable, but a truly general AI physicist should be free to inspect even those basic assumptions.
Being able to see through a bad inductive bias is one of the most important qualities of a good physicist, and the key to many of the great advances in physics.
Ptolemy preferred an Earth centered cosmos.
Copernicus saw through that bias, but himself was .. towards perfectly circular orbits, a bias that was only later seen through by Kepler.
So can a machine start to see through inductive biases?
Well, the first step is to see how far it can go with minimizing the supplied structure.
The leanest possible assumed mathematical representation and see what it can figure out itself from raw data.
A landmark example is AI Feynma.. methods rather than relying on just one neural network.
Given numerical data, a neural network first helps expose hidden structure.
For example, whether the relationship separates into simple pieces or possesses useful symmetries.
Dimensional analysis checks whether the units are consistent, while symbolic regression searches through explicit mathematical expressions until it finds a compact equation.
Now, symbolic regression is the key advance here because it tries to find the underlying mathematical form rather than shoehorning relationships into a prescribed formalism on a benchmark drawn from the Feynman Lectures in Physics.
This system recovered all 100 equations.
The fact that AI Feynman is a hybrid system tells us something.
We're unlikely to be able to duplicate everything that a physicist does with one architecture, but it may be possible for multiple complementary methods working in tandem.
AI Feynman, still required a good amount of hand-holding.
Besides handing it the problems to solve, we chose the variables, provided clean numerical data, and told it roughly what sort of mathematical object should count.. But more recent systems have tried to relax even those assumptions.
Symbolic pregression goes a step beyond ordinary equation finding.
Instead of being handed the right physical variables, it can start from, say, raw video.
Learn a set of hidden coordinates in which the motion bec.. regression to discover equations governing those coordinates.
The method explicitly favors low dimensional latent variables and simple dynamics by minimizing things like non-linearity, acceleration, and prediction error, and the final symbolic regression prefers simple equations.
So the method is an assumption free.
It effectively implements Occam’s razor, assuming that the most true descriptions are the most simple ones.
But once it's settled on what it believes to be the most parsim.. It effectively finds representation and the law simultaneously.
That's starting to sound more physicist like it, even if a physicist Silviu-Marian Udrescu and Max Tegmark in this case invented the system, fitted the data, and verified the output.
More recently, pixel2phys has pushed the same general goal using multimodal foundation model agents.
Starting from simulated and real scientific video, d different parts of the system propose candidate physical variables, derive equations, test whether those equations reproduce the observed dynamics, and revise them when they fail.
There's a peculiar trade off here.
The systems that have given the fewest explicit assumptions in the physics have mostly shown that they can rediscover laws that we already know, which is a nice validation of the promise of these approaches.
But it's not yet new physics currently.
In order to actually do useful new physics, we need to lay down a lot more scaffolding, build in some inductive bias.
In other words, we can make progress in pre-established directions, but not open up entirely new theoretical landscapes.
Before we ask whether I will ever be able to identify truly novel directions for research, there's another aspect of physics, well, of all science, that's traditionally been the province of humans, and which AI is also encroaching on actual experimental design.
Many of the great advances in quantum physics have come from bouncing photons around on tabletops, starting with the classic double slit experiment, but getting much more complicated since then.
But there are innumerable configurations of lasers, mirrors, beams, splitters and other optical paraphernalia th.. What if there are quantum secrets to be revealed in obscure optical setups that mere humans can't hope to stumble upon?
In 2016, a machine learning system called Melvin searched through arrangements of components for quantum optics experiments.
Melvin worked by turning experimental architecture into a combinatorial search problem, in which he could learn and reuse useful structures discovered in earlier searches.
That was enough to find unusual experimental configurations that human researchers had missed.
In 2025 Urania applied a more advanced version of this idea to find gravitational waves interferometers by navigating through a sort of universal interferometer space by various methods, including gradient based local searches.
The result was dozens of unconventional detector configurations outperforming an optimized human baseline.
So now we have machines that can find the best mathematical representations of a problem and the best architecture for the experiment.
But the core of the scientific process still seems to be in the hands of the scientist.
Humans are still the go to for formulating an interesting question and a hypothesis with its testable predictions, and of understanding the outcome of the resulting chain of experimentation, analysis, law derivation, etc., and of integrating these findings in the existing framework of the field and the corpus of all scientific knowledge.
A critical factor is that the process of scientific discovery requires multiple steps of highly varied capabilities, from imagination to computation to synthesis, and the ability to plan and direct the entire process.
This last part always felt like something a human would be needed for.
But even that's changing with the advent of transformers and foundation models like LLMs and of reasoning models and of AI agents.
Let's start with transformers.
Transformers are neural networks built around attention mechanism that lets the model decide which other parts of the input are most relevant to whatever it's processing right now.
Instead of handing information strictly one step at a time, attention can directly connect distant pieces of a sequence and equation with a definition many paragraphs earlier, or a piece of code with the variable it depends on.
The context.
Window sets how much material the model .. So larger context windows let it keep more of a paper, a derivation, a code base, a data set, or the experimental record simultaneously active.
For science, that matters because research problems often depend on long range connections across these elements, and transformers therefore make it possible for one model to integrate that scattered information into what feels like a coherent line of reasoning.
This capability really unlocked the modern AI era.
Start with the transformer based architecture, pre-training on enormous data sets, code bases, and in the case of some LLMS, the entire internet and then post train with usually reinforcement learning to refine behavior, you end up with a model that's able to hold extraordinarily rich contextual information, and then to implement whatever that context encodes that can include natural language processing, .. reaching new levels of subtle integration of varied datasets.
The reinforcement learning part is enormously effective.
For example, it enables pre-trained models to learn sequential control policies.
In one example, a neural network learned in simulation how to control magnetic coils in the TCB- Tokamak and then transferred that policy to the real fusion device and maintained a variety of plasma shapes.
Reinforcement learning is excellent when the task can be framed as repeated actions with a well-defined reward, but scientific research is usually less tidy.
On the surface, it seems humans are better at long horizon planning and synthesis, as well as finding highly creative shifts within the program.
One of our key abilities is that we are contemplating if we can constantly update and inspect our own thought processes.
A raw LLM can't really do that.
So what happens when we give one that ability?
The key is what we call inference time compute.
A raw LLM can be trained through reinforcement learning to think carefully, to do more internal iteration before producing an output.
And they can also be controlled by an external scaffold to iterate over their outputs.
Either way, these approaches allow.. and decide on new alternative steps in general to talk to themselves.
The result is a reasoning model.
If you then incorporate a reasoning model into a genetic system, give the m.. access to code, compilers, image and other data analysis tools, simulation tools, and all scientific papers ever written.
Then the results are frankly shocking.
The cosmic string result that we started this episode with is a.. The pre-trained reasoning model supplied broad mathematical knowledge and generated candidate approaches.
It's scaffold enabled tree search, which allowed the system to branch into multiple strategies, code and numerical tools checked intermediate results.
You can also give one of these systems access to physical experimental apparatus.
At synchrotron facilities, early versions of this have begun carrying out pieces of real experimental workflow, such as X-ray crystal alignment.
To recap, modern AI can perform many piecewise steps in a defined scientific project, from experimental design and execution to data analysis to law discovery, and can even iterate within steps and manage across multiple steps.
There are many teams working on end to end AI scientists, and a few on explicit AI physicists.
Maybe I'll talk about those another time, but for now, what's left for humans?
In a 2024 paper by Bettel, De and Gershman, they articulated what they call the easy and hard problems of AI for science.
The easy problem, which they admit is also hard, is using AI for these piecewise steps in the scientific process that I mentioned.
We could say that AI's are increasingly able to p.. or to correctly answer questions posed.
So that's the easy problem.
But they cannot do the hard problem, which.. is to come up with truly novel questions in the first place.
And by a novel question, I mean identifying a direction of inquiry that can lead to a profound shift in our understanding.
In a 2026 study called Astro Insight A Genetic Lens, systems read astronomy literature and generated research ideas that were then judged by human experts.
Some systems they tested were able to generate genuinely novel and promising research directions, but these ideas were mostly useful extensions and recombination of existing approaches.
There was no evidence of directions that would lead to true breakthroughs.
Modern AI is capable of what you might call combinatorial novelty.
It can absorb a field and reconfigure the parameters of past work to find what hasn't been done yet.
But the answers generated by such questions tend to remain firmly within the starting scientific framework.
That framework is the inductive bias in the scaffold that I currently needs in order to make scientific progress at all.
But for true.. and for that truly novel questions are needed.
Questions like Einstein's what if space and time and mutable and gravity and acceleration are equivalent?
Or Heisenberg's.
What if we describe only the observables of the atom?
In principle an AI could riff off past insights like these, but current AI seems to be too framework locked to identify genuinely novel questions.
Some describe modern AI is lacking scientific taste, which is a sort of instinct for novel questions and promising research paths.
Scientific taste requires the subtle integration of deep knowledge of fields and methods with curiosity, imagination, abstraction, anomaly, sensitivity, skepticism, even aesthetics, and more.
AI can do many of these things, but can't yet integrate them like a human can.
And integration gets me back to the one thing that currently only humans can do, and maybe the one thing will always be nee.. understanding what a new result means in the context of everything else we know about the world.
AI is unlikely to fully solve physics by itself, but even if it found the master equatio.. who cares if no one understands what it means?
The most capable AI physicist is pointless if its findings don't feed back into the integrated scientific worldview that humans have been building for millennia.
Physics is going to change due to AI.
It already has for bad as well as for good.
And at least for now, AI is best when directed by human physicists for the purpose of actual humans understanding the nature of space time.
Support for PBS provided by:
















