Korva is a scroll of the internet’s best ideas.
AI task length grows tenfold yearly
The tenfold pattern
Grade school math took a human mathematician about five seconds. The next year's benchmark took an expert about a minute, and the year after that AIME, the qualifier for the USA Mathematics Olympiad team, took a good mathematician about ten minutes. Each step was roughly ten times longer than the one before, and the models cleared each one a year later.
Olympiad gold
Gold at the 2025 International Math Olympiad fit the line, since an Olympiad problem takes a human mathematician about a hundred minutes. Carrying the line forward gives about fifteen hours of human work for the following year, which should not be enough to solve a Millennium Prize problem.
The forecast
On that trend, a Millennium Prize problem looked out of reach for 2026 and 2027, with 2028 the earliest realistic guess. An AI system solved one in 2026.
Why is the AIME a useful benchmark?
The American Invitational Mathematics Examination is a 15-question, three-hour contest whose answers are whole numbers from 0 to 999. That format lets a program grade a model's answer exactly, with no human reading proofs. Problems need several chained steps rather than recall, so they separate real reasoning from memorized formulas. Top scorers on the AIME qualify for the USA Mathematical Olympiad.
How does an Olympiad exam differ from the AIME?
The International Mathematical Olympiad has six problems spread over two days, with 4.5 hours each day. Contestants write full proofs, and human graders award up to seven points per problem, for a maximum of 42. A numerical guess earns nothing, so a model has to produce a correct argument. In 2025, the gold-medal cutoff was 35 points.
How do Millennium problems differ from Olympiad ones?
Olympiad problems are written so that a talented teenager can solve them in hours, and the organizers know a solution exists. Millennium problems carry no such guarantee and have resisted professional mathematicians for decades. The Poincaré conjecture, posed in 1904, was settled only when Grigori Perelman posted his proofs online in 2002 and 2003.
More from Dwarkesh Podcast
