How Is AI Changing Mathematical Discovery?

TL;DR

AI excels unevenly across mathematics, solving some benchmark problems rapidly while still struggling with tasks that demand playful reasoning, valuable conjectures, or new definitions. Its most consequential future role may be connecting ideas across specialized fields, but creating entirely new conceptual frameworks would signal a broader form of intelligence with effects far beyond mathematics.

Transcript

Today, I'm chatting with Grant Sanderson, who runs  3Blue1Brown and is now working on a new project   documenting the progress AI is making in math. I wanted to talk to you about this because AI   has been making the fastest progress  in mathematics out of any other field.  Whatever is happening here, and whatever way  we're seeing AI progress happ... Read More

Key Insights

• AI progress in mathematics is uneven rather than uniform. Strong results on one category of problems can coexist with serious weaknesses in another, creating a spiky frontier whose irregularity remains visible even when researchers examine narrower mathematical domains.
• International Math Olympiad performance is not a decisive test of general intelligence. Olympiad problems may appear to demand creativity, but many can be trained for, and a system's final score can depend heavily on which categories happen to appear on a particular test.
• Geometry has been unusually tractable for specialized AI methods. The discussed system could solve Olympiad geometry through brute-force techniques in nineteen seconds, while combinatorics remained a harder and more playful category that contributed to weaker overall performance.
• A famous conjecture can be solved through different kinds of intelligence. Connecting two mature fields requires broad knowledge and analogy, while building an entirely new mathematical framework requires the creation of concepts that current systems may not yet demonstrate.
• Cross-domain connection is a plausible strength of language models. A system with deep knowledge of quantum physics and analytic number theory could notice shared mathematical expressions without depending on specialists from separate disciplines happening to meet and exchange observations.
• The Riemann hypothesis illustrates why solution character matters. A proof based on connecting established domains would differ substantially from one requiring a new mountain of theory, even though both outcomes would resolve the same celebrated mathematical question.
• Mathematical benchmark goals are vulnerable to moving goalposts. Once AI achieves a capability previously considered impressive, attention shifts toward harder abilities, such as proposing worthwhile conjectures, inventing definitions, or determining which mathematical objects deserve study.
• Great mathematical work includes choosing what to investigate. Proving supplied theorems is only one level of contribution, while generating significant conjectures and creating productive definitions may better reveal whether a system can shape a field rather than merely solve assigned problems.

Questions & Answers

Q: Why does success at the Math Olympiad not prove AGI?

Success at the International Math Olympiad does not prove general intelligence because performance can depend on specialized methods, training, and the mixture of problem categories. A system may solve geometry almost automatically while struggling with combinatorics. A gold-level aggregate score can therefore hide narrow strengths and weaknesses, just as passing earlier benchmarks did not produce a single moment when broad intelligence became established.

Q: Why is AI progress in mathematics described as spiky?

AI progress is described as spiky because capability varies sharply across mathematical tasks. Even within a single competition, systems may be extremely effective at geometry but much less reliable at playful combinatorial problems. The same pattern continues when researchers zoom further into mathematics, so apparent strength in the field as a whole does not imply consistent reasoning across every specialty or problem type.

Q: How did problem selection affect AI performance at the IMO?

Problem selection mattered because the International Math Olympiad contains six problems drawn from geometry, number theory, algebra, and combinatorics. In 2024, two problems were in combinatorics, an area where the system struggled, while geometry was much easier for it. The discussion argues that a test containing more geometry questions could have produced a gold-level result, showing how category distribution influences the headline score.

Q: How could AI connect separate fields of mathematics?

AI could connect separate fields by recognizing that the same structure, expression, or statistical pattern appears in different bodies of knowledge. The conversation uses the link between zeros of the Riemann zeta function and eigenvalues of random Hermitian matrices as an example. A system knowledgeable across quantum physics and analytic number theory might notice such similarities without relying on a chance conversation between specialists.

Q: What would make an AI proof of the Riemann hypothesis especially significant?

The significance would depend on what the proof required. If AI connected concepts already developed in separate fields, the result would demonstrate powerful breadth and analogy. If it had to create a new mathematical framework, building the right concepts before the proof could even be formulated, that would indicate a more profound kind of intelligence that would likely influence domains beyond mathematical problem solving.

Q: What does Fermat's Last Theorem illustrate about mathematical discovery?

Fermat's Last Theorem illustrates that a simply stated problem may require a large and sophisticated conceptual structure. The eventual approach connected extensive bodies of ideas centered on elliptic curves and modular forms. This example distinguishes solving a problem with existing machinery from first developing the intellectual mountains needed to ask and answer the right connecting question.

Q: What should replace theorem solving as the next AI math benchmark?

Possible next benchmarks include generating interesting conjectures, identifying valuable problems, and creating definitions or objects that organize and unify fields. These tasks go beyond proving a theorem supplied by someone else. They test whether AI can determine what is worth studying and develop conceptual tools that guide future research, although their open-ended value makes them difficult to score with a clear pass-or-fail standard.

Q: Why are conjecture and definition generation difficult to benchmark?

Conjecture and definition generation are difficult to benchmark because there may be no immediate, objective goal line. A theorem can often be checked as proved or unproved, which supports clear evaluation and reinforcement from verifiable results. By contrast, the importance of a new question or definition depends on whether it reveals structure, inspires productive work, or eventually unifies ideas, and that value may emerge only later.

Summary & Key Takeaways

• AI progress in mathematics has a spiky, fractal structure. Systems can perform extremely well in certain areas, such as Olympiad geometry, yet struggle with combinatorics. Passing a prominent benchmark therefore does not establish general intelligence, because aggregate performance can conceal major differences among problem types and the methods used to solve them.

• The nature of a major mathematical solution matters more than the prestige of the problem alone. A system might solve a famous conjecture by connecting knowledge from established fields, or it might need to invent a new body of theory. The second achievement would represent a substantially different and more broadly transformative capability.

• Future evaluations may need to measure whether AI can identify worthwhile questions, formulate illuminating conjectures, and invent definitions that organize or unify mathematical fields. These abilities are difficult to benchmark because their value may not be immediately verifiable, but they better reflect the conceptual work associated with the greatest mathematical contributions.