GPT-5.4 Pro proof method cracks more Erdős problems

The interesting part here is not just that the model hit one hard problem, but that its proof structure appears reusable enough to unlock adjacent results. That is a stronger signal for AI-assisted mathematics than a single isolated solve.

• –The tweet frames this as one of the first downstream impacts of an AI-generated proof, which is the real headline for research workflows
• –Reuse across multiple problems suggests the model surfaced a useful proof method, not just a lucky answer
• –The arXiv follow-up matters because it turns a viral claim into something the math community can inspect and build on
• –If this holds up, the value of frontier models in math may be less about one-shot theorem solving and more about proposing novel proof templates