ChatGPT Quietly Cracks Open Math Problems

[johnblockbuster]

Over the weekend, software engineer and former quant researcher Neel Somani discovered something unexpected while testing OpenAI’s newest model: ChatGPT produced a correct, end-to-end solution to an open math problem after “thinking” for about 15 minutes. When Somani checked the reasoning and formalized it using the Harmonic toolchain, the proof held up.

What stood out wasn’t just correctness, but depth. The model referenced advanced concepts like Legendre’s formula and Bertrand’s postulate, eventually triangulating toward earlier academic work — including a 2013 MathOverflow solution by Noam Elkies. Still, ChatGPT’s final proof differed in key ways and addressed a version of the problem originally posed by Paul Erdős, whose unsolved conjectures have become a benchmark for AI-assisted mathematics.

[JanEmil]

Fantastic and scary the same time.

[ed]

feels less like “ai replaces mathematicians” and more like “ai never gets tired”