AI and the Future of Mathematics Part 2: What LLMs Can and Can’t Do
In Part 1 of this series, I explored how the advent of computers transformed the field of math. Now, LLMs are ushering in another revolution, and their capabilities are at an inflection point.
Last year, I had a side gig teaching AI how to do math. I was paid to create original problems that stumped frontier models and provide solutions for future models to train on.
The Bad
During my stint, I identified several fundamental issues that challenged AI. For example, if a problem required a TON of computation with many decimal places, it might round digits in odd places and be off by a small fraction. Large computational questions were time-intensive to write, so I focused on one of my favorite subjects: algebraic topology. I realized that AI didn’t fully understand how shapes exist in space…
Consider the following problem:

Draw three equally sized circles on a plane such that each pair of circles touches at exactly one point. Remove the circles. How many sections are there?
In math terms: Consider the following circles as subsets of ℝ²:
Let C = C₁ ∪ C₂ ∪ C₃. How many path-connected components does ℝ² ∖ C have?
There’s a space inside each circle, an area outside of the circles, and a smaller section between all three of them. So the answer is 5. AI answered 4. It couldn’t intuit the inner section between the circles. I exploited this lack of geometric understanding to derive many questions that AI couldn’t solve.
The Good
What was surprising, however, was how good LLMs were at most math. I could ask graduate-level algebra questions, and if the answer required a proof or computation using a strategy that had been documented, the model would solve it. It seems that LLMs lack the creativity to derive solutions from methods outside their training data. But not many problems need fundamentally novel ideas to solve…
So what else is AI good at?
- AI is exceptional at finding counterexamples. It is unbiased, tenacious, and world-class at utilizing computation to verify statements. LLMs can generate examples and analyze them much faster than mathematicians can.
For example, ChatGPT’s model found a configuration of dots that increases the number of pairs exactly one unit apart, exceeding the conjectured limit and disproving Erdős’s Problem 90. Additionally, Anthropic’s model disproved the Jacobian Conjecture in three dimensions by finding a polynomial map with a constant nonzero Jacobian determinant that is not invertible.
- AI models can conceptualize all of math and utilize ideas across the discipline. Even the most studied mathematicians are limited in their global understanding of math—AI has no such limitation.
LLMs recently proved the quasi-Riemann hypothesis by combining ideas from disparate fields: cubic Gauss sums, metaplectic theta functions, and large-sieve estimates.


Epilogue
Several months after the math dataset project, I was invited back to participate in Round 2. It became clear the frontier models had trained on my questions and solutions. The models were considerably smarter, and I couldn’t create one problem that challenged them.
It’s worth a moment to reflect on this timeline: I stumped frontier models with basic problems like the one above in 2025. One year later, generative AI solved the Navier–Stokes equation by producing an explicit vector field that blows up in finite time.
I’ll never be able to tell my kids I solved a Millennium Prize Problem. But I can tell them I tutored the machines that did.
Discover more from Cooper Young
Subscribe to get the latest posts sent to your email.