AI solves math problem that stumped experts for years
GPT-5.4 just cracked a math challenge that was designed over nearly 20 years to challenge AI systems.
The problem, created by Dr. Bartosz Naskrecki and designed to be nearly impossible, even for Ph.D.s, was finally solved by AI after nearly 20 years of effort went into making it.
Problem designed to be extremely difficult for both humans, models
Naskrecki's puzzle blends ideas from several advanced math fields and fills 13 dense pages.
Even top mathematicians would need at least a month to try solving it, but GPT-5.4 produced a correct solution in testing (after multiple attempts); the problem had been designed to be extremely difficult for both humans and models.
Solution 'almost human'
Before GPT-5.4, the hardest group of FrontierMath problems had only seen four solutions from any AI model; GPT-5.4 produced a correct solution to one of those problems.
Naskrecki called its solution "almost human," showing just how quickly AI is catching up with real mathematicians.