AI starts to take power of discovering things from mathematicians.
On October 6, 2026, OpenAI published a remarkable collection of mathematical research.
372 families of mathematical results.
More than 700 research manuscripts.
Around 4,000 mathematical problems evaluated.
Claimed solutions to problems that have challenged mathematicians for generations.
Some results include machine-checked proofs. Others still require independent verification. These aren’t all confirmed breakthroughs, and several manuscripts have required revisions.
But even with those qualifications, something important is happening.
For centuries, the relationship between humans and tech has been relatively simple.
Humans discover. Machines execute.
Humans discovered electricity. Machines helped us use it.
Humans developed quantum mechanics. Engineers built semiconductors.
Humans created mathematical theories. Computers performed calculations based on them.
Now AI is beginning to participate in the discovery process itself.
And that’s a completely different economic and power proposition.
I want to explain 25 mathematical problems from OpenAI’s research, what they mean in plain English, and why they could eventually matter to science, industry, and ordinary people.
Let’s begin.
I – You’re looking at the wrong part of the value chain
AI applications today do one of three things:
Make existing work faster.
Make existing work cheaper.
Automate work that humans previously performed.
Useful? Absolutely.
Transformative? Sometimes.
But there’s another layer.
Think about your smartphone.
It exists because generations of scientists discovered things about electricity, electromagnetism, quantum physics, and mathematics.
The smartphone didn’t create those discoveries.
It commercialized them.
The same is true for aviation, medical imaging, GPS, encryption, and almost every major technology.
Discovery creates possibilities. Engineering turns them into products. Business distributes those products to the world.
If AI can accelerate discovery, it could affect the entire chain.
That’s what makes these mathematical results interesting.
II – 25 mathematical problems and why they matter
You don’t need to understand the equations.
You need to understand what scientists are trying to discover.
1. Quasi-Riemann Hypothesis
Prime numbers appear unpredictably among ordinary numbers. Mathematicians have spent generations trying to understand their distribution.
OpenAI claims a major result restricting where zeros of certain mathematical functions can occur.
Why it matters: Prime-number theory helps underpin modern cryptography. Deeper understanding could eventually improve mathematical algorithms. It doesn’t solve the full Riemann Hypothesis or break encryption.
2. Artin’s Primitive Root Conjecture
Can certain numbers generate every possible nonzero remainder through repeated multiplication?
OpenAI claims the answer is yes for infinitely many suitable primes.
Why it matters: This mathematics is connected to modular arithmetic, which appears in cryptographic systems.
3. Erdős–Prachar Large Prime Gaps
How frequently do unusually large gaps appear between prime numbers?
OpenAI claims a strong mathematical result about their frequency.
Why it matters: Understanding prime distribution improves the mathematical foundations of number theory and related algorithms.
4. Erdős–Hall Totient Counting Problem
How often do different numbers produce the same kind of arithmetic counting result?
OpenAI claims a new asymptotic formula.
Why it matters: Euler’s totient function appears in the mathematics behind RSA encryption, although this result does not directly change its security.
5. Patterson’s Cubic Gauss Sum Conjecture
Can we predict the average behavior of complicated arithmetic sums?
OpenAI claims a precise formula.
Why it matters: It gives mathematicians stronger tools for understanding patterns in numbers.
The takeaway?
We’re still discovering fundamental properties of numbers that we’ve been studying for thousands of years.
2. The limits of computers and intelligence
This is where things get interesting.
Because some mathematical problems aren’t about finding an answer.
They’re about whether finding an answer is even possible.
6. Hilbert’s Tenth Problem Over Rational Numbers
Can a computer always determine whether a polynomial equation has a solution made entirely of fractions?
OpenAI claims to have proven that no universal algorithm can do this.
Why it matters: It establishes a fundamental limit on computation. Even infinitely powerful computers cannot solve every mathematically defined decision problem.
7. Kaplansky’s Direct-Finiteness Conjecture
In certain abstract mathematical systems, does multiplying A by B to get 1 mean multiplying B by A must also give 1?
OpenAI claims a counterexample.
Why it matters: It changes our understanding of certain algebraic structures used in advanced mathematics.
8. Erdős’s Short Egyptian Fractions Conjecture
How efficiently can numbers be represented as sums of distinct fractions with numerator one?
OpenAI claims a near-optimal bound.
Why it matters: It advances the mathematics of numerical representation, though no immediate compression or software application is demonstrated.
9. Irrationality Exponent of Pi
We know π never ends or repeats.
But how closely can it be approximated using fractions?
OpenAI claims to have established its exact irrationality exponent.
Why it matters: It resolves a fundamental question in mathematical approximation.
10. Irrationality of Catalan’s Constant
For generations, mathematicians couldn’t prove whether the constant 0.915965... could be expressed as a fraction.
OpenAI claims it cannot.
Why it matters: Primarily foundational mathematics, but the techniques may help solve other problems.
Notice something?
Not every discovery creates an immediate product.
Some expand our understanding of what’s mathematically possible.
3. Physics, energy, and materials
Now we’re getting closer to the physical world.
11. Quantum Heisenberg Ferromagnet
Why do collections of quantum particles spontaneously produce magnetism?
OpenAI claims rigorous results for certain mathematical models.
Why it matters: Could strengthen research into magnetic materials, memory technologies, and spintronics.
12. Mézard–Parisi Formula
How do large systems behave when millions of interacting elements influence one another?
OpenAI claims a proof for certain spin-glass models.
Why it matters: Related mathematics appears in statistical physics and theoretical optimization, potentially influencing future computational methods.
13. Relativistic Vlasov–Maxwell System
Can we mathematically predict the behavior of charged particles interacting with electromagnetic fields without the solutions developing singularities?
OpenAI claims a global mathematical result for a specified model.
Why it matters: These equations are related to plasma physics, an important field in fusion-energy research.
Better mathematics could eventually improve scientific modelling. It doesn’t automatically produce better fusion reactors.
14. Higher-Dimensional Duke Equidistribution
How do certain mathematical objects spread across complicated, higher-dimensional spaces?
OpenAI claims a significant equidistribution result.
Why it matters: Advances our understanding of number theory and mathematical dynamics.
15. Mahler’s Geometric Conjectures
Which shapes minimize a particular geometric measure, even in hundreds of dimensions?
OpenAI claims solutions to important versions of these conjectures.
Why it matters: High-dimensional geometry is connected to optimization and theoretical computer science.
Here’s the important distinction.
A mathematical discovery is not an engineering breakthrough. But engineering breakthroughs often depend on mathematical discoveries.
4. The hidden structures of mathematics
Sometimes the hardest problems aren’t about numbers.
They’re about discovering relationships between things that appear unrelated.
16. Milne’s Rationality Conjecture
Can different mathematical descriptions of certain complex geometric objects be proven consistent?
OpenAI claims a proof.
Why it matters: Strengthens foundations in advanced geometry and number theory.
17. Birch–Swinnerton-Dyer
Can we predict the structure of rational solutions to certain elliptic-curve equations?
OpenAI claims a major result for a specified class of curves.
Why it matters: Elliptic curves are central to number theory and have important cryptographic applications. The full conjecture remains a separate question.
18. Goldfeld’s Conjecture
How frequently do families of elliptic curves have particular numbers of independent rational solutions?
OpenAI claims a result about their distribution.
Why it matters: Improves our understanding of mathematical structures used in advanced number theory.
19. Zilber–Pink Conjecture
How often can unusually special intersections occur between complicated geometric objects?
OpenAI claims important results for specified cases.
Why it matters: Advances algebraic geometry and the study of exceptional mathematical patterns.
20. Function-Field Reconstruction
Can you reconstruct a complicated mathematical structure using limited algebraic information?
OpenAI claims new reconstruction results.
Why it matters: Helps mathematicians understand how much information determines an entire structure.
The bigger lesson?
Sometimes progress comes from discovering that two apparently different problems share the same underlying structure.
That can unlock completely new ways of solving them.
5. The mysteries hidden inside simple mathematical ideas
21. Chowla and Elliott Correlation Conjectures
Do patterns involving prime factors behave randomly, or do hidden relationships exist?
OpenAI claims important correlation results.
Why it matters: Advances mathematical understanding of arithmetic randomness.
22. Weak Inhomogeneous Duffin–Schaeffer Conjecture
How closely can numbers be approximated using fractions under specific constraints?
OpenAI claims a proof establishing infinitely many approximations under the conjecture’s conditions.
Why it matters: Strengthens foundational approximation theory.
23. Gaussian Moat Conjecture
Imagine prime numbers scattered across a two-dimensional grid.
Can you travel forever from one prime to another if every jump must remain shorter than a fixed distance?
OpenAI claims no.
Why it matters: Advances mathematical understanding of connectivity and prime-number geometry.
24. Deligne–Drinfeld Conjecture
Can a complicated mathematical structure be explained using a surprisingly simple collection of components?
OpenAI claims a proof of the relevant algebraic statement.
Why it matters: Simplifying mathematical structures can unlock further discoveries.
25. Litt’s Integral Density Question
How common are certain special arithmetic points inside complex geometric spaces?
OpenAI claims an important density result.
Why it matters: Strengthens the relationship between arithmetic and algebraic geometry.
III – The biggest mistake is expecting these discoveries to make your phone faster tomorrow
Think about the history of electricity.
Long before the modern electrical grid existed, scientists were trying to understand electromagnetic phenomena.
They weren’t developing power stations or charging infrastructure.
They were discovering how nature worked.
The commercial applications came later.
The same thing happened with quantum mechanics.
Scientists weren’t trying to invent smartphones.
They were investigating the behavior of matter.
Decades of research, experimentation, and engineering eventually helped make semiconductors possible.
That’s how much of scientific progress works.
Mathematics → Scientific understanding → Engineering → Industry → Consumer benefit.
And not every discovery completes that journey.
Some remain theoretical forever.
But when a foundational discovery does become useful, its economic impact can be enormous.
That’s why focusing only on today’s AI productivity gains may be missing a much bigger opportunity.
IV – What happens when the cost of discovery collapses?
Historically, scientific progress has depended on scarce resources.
Human intelligence.
Specialized expertise.
Research funding.
Time.
Sometimes a mathematical problem consumes the careers of multiple researchers without being solved.
Now imagine a future in which AI systems can generate, test, and formally verify substantial mathematical results far more efficiently.
What changes?
First, the economics of research.
A laboratory might explore significantly more hypotheses with the same budget.
A university could investigate mathematical ideas that previously required years of specialized work.
A company might accelerate parts of its R&D pipeline.
But here’s where things become more interesting.
If intelligence becomes abundant, intelligence itself may no longer be where the greatest economic value sits.
Imagine AI helps discover a mathematical principle that eventually improves battery chemistry.
Who captures the value?
The company operating the AI model?
The scientist who designs the experiment?
The manufacturer producing the battery?
The business controlling distribution?
Or the consumer buying cheaper energy storage?
Probably some combination.
The point is that scientific discovery and commercial value creation are not the same thing.
The scarce resources may shift toward physical infrastructure, experimental facilities, capital, manufacturing, proprietary data, and distribution.
A theorem can become abundant.
Building a factory cannot.
At least, not in the same way.
For years, the AI race has been dominated by a familiar set of questions.
Who has the best model?
Who has the most GPUs?
Who owns the data centers?
Who can deliver intelligence at the lowest cost?
Those questions matter.
But I think another competition is emerging.
Who can turn intelligence into discoveries that change the physical economy?
Consider two businesses.
One uses AI to reduce customer service costs by 50%.
The other uses AI-assisted research to discover a new industrial process that cuts the cost of producing an important material by 50%.
Assuming the second discovery works and can be commercialized, the economic implications are very different.
The first business has improved an existing process.
The second may have changed the economics of an entire industry.
That’s the difference between making existing knowledge cheaper and creating new knowledge that enables something previously impossible.
And it’s the reason I think AI-driven scientific discovery deserves much more attention.
VI – The most valuable thing AI could create isn’t content. It’s knowledge.
Let’s return to those 25 mathematical problems.
Some involve prime numbers.
Some involve quantum physics.
Some concern the mathematical limits of computation.
Others deal with geometric structures so abstract that most people will never need to understand them.
But they have something in common.
They represent attempts to answer questions humanity hasn’t previously been able to resolve.
OpenAI’s claimed results still need rigorous verification. Publishing hundreds of papers is not the same as establishing hundreds of accepted mathematical breakthroughs.
But if a meaningful portion survives scrutiny, it tells us something important.
The role of AI is expanding.
From performing tasks.
To helping humans reason.
To potentially producing original knowledge.
And knowledge is what ultimately makes new technologies possible.
We spend an extraordinary amount of time debating whether AI will replace human jobs.
That’s a legitimate concern.
But perhaps the more consequential question is what happens when AI helps humanity discover things we couldn’t have discovered otherwise.
New mathematical foundations.
New physical models.
Potentially new materials, medical treatments, and energy technologies after the necessary experimental and engineering work.
The greatest economic impact of AI may not come from replacing the work we already do.
It may come from enabling discoveries that create entirely new industries.
For most of history, our technological capabilities have been limited partly by how quickly we could expand human knowledge.
What happens when that constraint starts to loosen?
We don’t know yet.
But it might turn out to be a much bigger story than another chatbot replacing another job.