🧠 GPT-5.4 just “solved” a 20-year-old math problem… by finding a forgotten paper
A new story circulating in AI circles claims GPT-5.4 cracked a decades-old math problem from the FrontierMath benchmark, a set of research-level problems designed to challenge top mathematicians. But the twist makes the story even more interesting.
Instead of inventing a completely new proof, the model found an obscure 2011 preprint paper that already contained the key idea. The benchmark author didn’t know the paper existed, so the problem had been considered unsolved.
What happened:
• Researchers tested GPT-
A new story circulating in AI circles claims GPT-5.4 cracked a decades-old math problem from the FrontierMath benchmark, a set of research-level problems designed to challenge top mathematicians. But the twist makes the story even more interesting.
Instead of inventing a completely new proof, the model found an obscure 2011 preprint paper that already contained the key idea. The benchmark author didn’t know the paper existed, so the problem had been considered unsolved.
What happened:
• Researchers tested GPT-
