AI Programs From OpenAI and Google Score Gold-Medal Level at the World's Hardest Math Olympiad
Both companies said their systems solved five of six problems at the International Mathematical Olympiad, the elite competition for high-school mathematicians, reaching the gold-medal threshold under the same time limits as the students.
OpenAI and Google DeepMind both announced today that their AI systems reached gold-medal performance at this year's International Mathematical Olympiad, the most prestigious math competition in the world, where a few hundred of the best high-school mathematicians from around the globe sit two four-and-a-half-hour exams. Each company's system solved five of the six problems, scoring 35 of 42 points — above the gold-medal threshold. The problems were the real ones from this year's competition, and the systems worked under the same time limits, writing out proofs in plain language.
This is a milestone the field has been chasing and dreading in equal measure. Olympiad problems are not calculation; they require the kind of creative insight — finding the one clever step that unlocks a proof — that has always been held up as distinctly human, and that AI was supposed to be bad at. Three years ago these programs struggled with arithmetic. Google reached silver-medal level last year with a system that needed problems translated into a special format; this year's systems worked directly from the English, using general reasoning.
There are caveats. Google worked with the competition's organizers and had its result officially graded; OpenAI graded its own and announced ahead of the official process, which annoyed some. The systems are experimental and not available to the public. And solving contest problems is not the same as doing original mathematics, which is open-ended in a way a timed exam is not.
Still, the direction is what unsettles even the researchers who built these systems. A skill that was supposed to be a decade away, and to mark something special about human minds, arrived this summer, in two labs at once.