KorvaJoin the waitlist
Tech

AI Went From Botching Addition to Solving Decades-Old Math Problems

Because math answers can be checked, systems get the practice they need to improve fast.

AI image
Dwarkesh Podcast · The avalanche that hit math is coming for AI development

Three years ago, artificial intelligence could not reliably add four-digit numbers. Now it is solving math problems that top mathematicians and computer scientists chased for decades.

In 2024, systems could barely solve problems from a high school math contest. A year later, they performed at the gold medal level of the International Math Olympiad, the elite competition for high school students.

Since then, models have cracked niche open problems, with hundreds of long sought results appearing as side projects around new releases. About a month ago, an OpenAI system solved a Millennium Prize problem, among the most famous hard problems in math, in 88 hours.

Math gives AI something to practice against: answers that can be checked. Systems are strongest at verifiable work like finding counterexamples, rather than inventing broad new theories.

Questions people ask

What is a Millennium Prize problem?

They are seven famously hard math problems selected in 2000, each with a one million dollar prize for a correct solution. Only one has been fully solved by humans. In this context, they serve as a benchmark for solving problems that have resisted experts for years.

How hard is the International Math Olympiad?

It is an annual elite mathematics competition for high school students from more than 100 countries. Each student tackles six very hard problems over two days. A gold medal performance means solving at a level reached by only a small group of the world's strongest young mathematicians.

Why does having checkable answers help AI improve?

In math, a proposed proof or counterexample can be checked step by step, so a system gets clear right or wrong feedback. That feedback allows extensive trial and error. In more open ended subjects, good answers are harder to define and score automatically.

More in Tech