How AI Is Upending the World of Mathematics

PodcastOdd LotsNoah GiansiracusaOct 9

AI is changing mathematical research by helping complete proofs, while raising questions about verification, credit, research incentives, and who controls access to increasingly centralized tools. The guest says proof production is becoming less of a constraint than human review, and that current systems generally recombine existing knowledge rather than generate entirely new theories. The discussion also considers classroom adaptation and argues that mathematical judgment—especially choosing meaningful questions—remains important.

Artificial Intelligence AI Safety Technology Stocks

Preview · 27 more points in the app

Read the Full Summary

Key Takeaways

The hosts describe mathematics as a benchmark AI companies use to support claims about reasoning and AGI, while the guest emphasizes that mathematics also spans practical applications and work pursued for understanding or intellectual enjoyment. Examples range from abstract areas that later informed modern cryptography to the equations and algorithms needed to simulate movie effects.

Mathematical verification matters both for AI-generated proofs and for simulations whose outputs people rely on: Lean lets software check proofs, while verified mathematical results can help establish confidence in simulation algorithms. The guest points to earthquake-related building-collapse assessments as a consequential use of finite element modeling, while noting that users must still trust the software itself.

The guest says AI may help complete long-standing mathematical problems, but describes the Navier–Stokes work as a human–AI effort: mathematicians developed much of the construction, and AI contributed critical insights and helped with the final steps. He also says the proposed proof had not been officially verified and characterizes its correctness and the claim that OpenAI was first to meet the criteria tentatively.

The guest describes a shift toward proof abundance, with human verification and peer review struggling to keep pace: ICLR submissions grew from roughly 1,000–2,000 about a decade earlier to 60,000, and reviewers may lack the bandwidth for thorough checks. He also says traditional indicators of mathematical research quality—including publication venue, volume, interest, and prizes—have become less aligned, complicating evaluation.

AI-assisted proofs raise questions about both credit and provenance: the guest questions how much intellectual contribution comes from prompting a model to address a major conjecture, and says this complicates judgments about human achievement and hiring. He also calls AI data attribution a critical problem, saying current tools are poor at determining whether generated material draws on existing work.

TECHNOLOGY

AI Benchmarks — The hosts say AI companies treat mathematics as a kind of holy grail for demonstrating that models can reason and supporting claims that they have achieved AGI.

“Math is an interesting one because you can tell that a lot of the AI companies themselves have an interest in math. Right?…”—

Applied Mathematics — The guest distinguishes mathematics pursued for practical applications from work pursued for intellectual enjoyment or to understand the universe. Modern cryptography shows how practical developments can emerge from areas once considered highly abstract.

“But essentially, some people go into math because they want to see it translated into applications or to bring fundamental insight into the kind of things we do every day…”—

Movie Simulations — The guest explains that simulating movie effects—including cloth, smoke, fluids, and rigid bodies—requires algorithms for partial differential equations. Making the simulations work well enough that artists do not have to think about the math is a challenging mathematical problem.

“So when you watch a typical movie with different movie effects, there's cloth, there's smoke, there's all kinds of fluids, all, you know, rigid bodies flying around…”—

AI And Mathematical Research

The hosts introduce concerns about AI's growing mathematical ability, the future role of mathematicians, and how students and researchers may need to adapt.

Proofs, Problems, And Purpose

Solomon distinguishes proving facts from choosing meaningful questions, and describes mathematics as a social pursuit driven by applications, curiosity, and collaboration.

Applied Math In The Real World

Applied mathematics models practical systems such as weather and movie effects, while pure and applied work often overlap as abstract ideas find later uses.

Navier–Stokes Explained

Solomon explains the fluid dynamics equations and the open existence-and-uniqueness question, then outlines how a counterexample could show singular behavior.

Lean Proof Verification

AI can propose proofs, but systems such as Lean translate formal claims into code that independent software can check against mathematical rules.

Why Proofs Need Trust

Solomon argues that verified mathematics matters in engineering because simulations and calculations can inform consequential decisions, including whether structures withstand earthquakes.

Mathematics And Research Incentives

AI challenges how mathematicians choose problems, measure productivity, evaluate researchers, and assign value to proof production versus human insight.

Proof Abundance Overloads Review

Solomon describes AI-generated claims reaching academics and warns that a flood of submissions is overwhelming peer review and weakening familiar quality signals.

P Versus NP And AI

The conversation explains P versus NP and revisits the Navier–Stokes counterexample as a human-and-AI effort that builds on earlier mathematical work.

Credit For AI Proofs

Solomon says academia lacks clear rules for credit when humans formulate problems and AI contributes key proof steps, raising questions about incentives and attribution.

Access To Powerful Models

The speakers worry that expensive compute and unreleased models give well-funded labs advantages over university researchers and students.

Elegance Versus Brute Force

Solomon says the Navier–Stokes result may be correct without being elegant, and warns that frictionless proof production can remove the insight gained along the way.

AI And New Mathematical Ideas

Solomon says current systems mostly recombine existing knowledge rather than produce entirely new theories, while data attribution remains unresolved.

Teaching Math In The AI Era

Solomon describes shifting education toward hands-on, social learning and assessment that tests students' understanding rather than AI's ability to complete assignments.